wehub-resource-sync
426e9eeabd
Benchmark Bridge Tests / benchmark (bunx @biomejs/biome check packages/lifeops-bench/src, benchmark-lint) (push) Waiting to run
Benchmark Bridge Tests / benchmark (bunx vitest run --config packages/lifeops-bench/vitest.config.ts --root packages/lifeops-bench --passWithNoTests, benchmark-tests) (push) Waiting to run
Build Agent Image / build-and-push (push) Waiting to run
Chat shell gestures / Chat shell gesture + parity e2e (push) Waiting to run
ci / test (push) Waiting to run
ci / lint-and-format (push) Waiting to run
ci / build (push) Waiting to run
ci / dev-startup (push) Waiting to run
Cloud Gateway Discord / Test (push) Waiting to run
Cloud Gateway Webhook / Test (push) Waiting to run
Cloud Tests / lint-and-types (push) Waiting to run
Cloud Tests / unit-tests (push) Waiting to run
Cloud Tests / integration-tests (push) Waiting to run
Cloud Tests / e2e-tests (push) Blocked by required conditions
CodeQL Advanced / Analyze (javascript-typescript) (push) Waiting to run
Deploy Apps Worker (Product 2) / Determine environment (push) Waiting to run
Deploy Apps Worker (Product 2) / Deploy apps worker to apps-control host (${{ needs.determine-env.outputs.environment }}) (push) Blocked by required conditions
Deploy Eliza Provisioning Worker / Determine environment (push) Waiting to run
Deploy Eliza Provisioning Worker / Deploy worker to Hetzner host (${{ needs.determine-env.outputs.environment }} @ ${{ needs.determine-env.outputs.deployment_sha }}) (push) Blocked by required conditions
Dev Smoke / Classify changed paths (push) Waiting to run
Dev Smoke / bun run dev onboarding chat (push) Blocked by required conditions
Dev Smoke / Vite HMR dependency-level smoke (push) Blocked by required conditions
Electrobun Submodule Guard / electrobun gitlink is fetchable (push) Waiting to run
gitleaks / gitleaks (push) Waiting to run
Markdown Links / Relative Markdown Links (push) Waiting to run
Publish @elizaos/example-code / check_npm (push) Waiting to run
Publish @elizaos/example-code / publish_npm (push) Blocked by required conditions
Publish @elizaos/plugin-elizacloud / verify_version (push) Waiting to run
Publish @elizaos/plugin-elizacloud / publish_npm (push) Blocked by required conditions
Quality (Extended) / Homepage Build (PR smoke) (push) Waiting to run
Quality (Extended) / Comment-only diff guard (push) Waiting to run
Quality (Extended) / Format + Type Safety Ratchet (push) Waiting to run
Quality (Extended) / Develop Gate (secret scan + UI determinism) (push) Waiting to run
Quality (Extended) / Develop Gate (lint) (push) Waiting to run
Sandbox Live Smoke / Sandbox live smoke (push) Waiting to run
Snap Build & Test / Build Snap (amd64) (push) Waiting to run
Snap Build & Test / Build Snap (arm64) (push) Waiting to run
supply-chain / sbom (push) Waiting to run
supply-chain / vulnerability-scan (push) Waiting to run
Build, Push & Deploy to Phala Cloud / build-and-push (push) Waiting to run
Test Packaging / Validate Packaging Configs (push) Waiting to run
Test Packaging / PyPI on Python ${{ matrix.python }} (push) Waiting to run
Test Packaging / Pack & Test JS Tarballs (push) Waiting to run
Test Packaging / elizaos CLI global-install smoke (node + bun) (push) Waiting to run
UI Fixture E2E / ui-fixture-e2e (push) Waiting to run
UI Fixture E2E / fixture-e2e (push) Waiting to run
UI Story Gate / story-gate (push) Waiting to run
vault-ci / test (macos-latest) (push) Waiting to run
vault-ci / test (ubuntu-latest) (push) Waiting to run
vault-ci / test (windows-latest) (push) Waiting to run
vault-ci / app-core wiring tests (push) Waiting to run
verify-patches / verify patches/CHECKSUMS.sha256 (push) Waiting to run
Voice Benchmark Smoke / voice-emotion fixture smoke (push) Waiting to run
Voice Benchmark Smoke / voiceagentbench fixture smoke (push) Waiting to run
Voice Benchmark Smoke / voicebench-quality unit smoke (push) Waiting to run
Voice Benchmark Smoke / voicebench TypeScript unit (no audio) (push) Waiting to run
Voice Benchmark Smoke / voice bench smoke summary (push) Blocked by required conditions
Windows CI / windows ([bun run --cwd packages/app-core test bun run --cwd packages/elizaos test bun run --cwd packages/cloud/shared test], app-and-cli) (push) Waiting to run
Windows CI / windows ([bun run --cwd packages/scenario-runner test bun run --cwd packages/vault test bun run --cwd packages/security test bun run --cwd plugins/plugin-coding-tools test], framework-packages) (push) Waiting to run
Windows CI / windows ([bun run --cwd plugins/plugin-elizacloud test bun run --cwd plugins/plugin-discord test bun run --cwd plugins/plugin-anthropic test bun run --cwd plugins/plugin-openai test bun run --cwd plugins/plugin-app-control test bun run --cwd plugins/pl… (push) Waiting to run
Windows CI / windows ([node packages/scripts/run-turbo.mjs run build --filter=@elizaos/core --filter=@elizaos/shared --filter=@elizaos/agent --concurrency=4 node packages/scripts/run-bash-linux-only.mjs scripts/verify-riscv64-buildpaths.sh node packages/scripts/run… (push) Waiting to run
Windows CI / windows ([node packages/scripts/run-turbo.mjs run typecheck --filter=@elizaos/core --filter=@elizaos/shared --filter=@elizaos/cloud-shared --concurrency=4 bun run --cwd packages/core test bun run --cwd packages/shared test], core-runtime, 75) (push) Waiting to run
Test Packaging / Build & Test PyPI Package (push) Waiting to run
Voice Workbench / headless workbench (mocked backends) (push) Has been cancelled
Voice Workbench / real acoustic lane (nightly, provisioned only) (push) Has been cancelled
1870 lines
60 KiB
YAML
1870 lines
60 KiB
YAML
# Source datasets to download and normalize.
|
||
#
|
||
# Fields:
|
||
# slug: short filesystem-safe name (used as data/raw/<slug>/ dir)
|
||
# repo_id: HF dataset repo id
|
||
# normalizer: adapter family in scripts/normalize.py
|
||
# weight: relative sampling weight in pack_dataset.py
|
||
# est_size_gb: rough on-disk estimate so the downloader can budget
|
||
# priority: "core" | "extra"
|
||
# license: SPDX-like license tag (required when consent_basis: licensed)
|
||
# consent_basis: REQUIRED. One of:
|
||
# synthetic | public_domain | licensed |
|
||
# opt_in_user_consent | internal_dogfood
|
||
# consent_proof_uri: REQUIRED when consent_basis is opt_in_user_consent
|
||
# or internal_dogfood. URI/URL of consent record.
|
||
# notes: free text
|
||
#
|
||
# The dataset_loader.py SOC2 consent gate enforces these fields at load time
|
||
# (SOC2 PI1.1-PI1.5, C1.1). Sources without consent_basis are rejected.
|
||
|
||
datasets:
|
||
# =================== CANONICAL TARGET FORMAT ===================
|
||
# ScamBench already publishes formats/eliza-train.jsonl — passthrough.
|
||
- slug: scambench
|
||
repo_id: shaw/scambench-training
|
||
normalizer: scambench_passthrough
|
||
weight: 4.0
|
||
est_size_gb: 0.5
|
||
priority: core
|
||
license: cc-by-sa-4.0
|
||
consent_basis: licensed
|
||
notes: |
|
||
Canonical eliza format already published as a config. We pull the
|
||
`eliza` config splits directly: formats/eliza-{train,test,validation}.jsonl
|
||
|
||
# =================== CLAUDE DISTILLATION (REASONING SFT) ===================
|
||
# Kassadin88/Claude-Distills — ~129k {system,user,assistant} chats
|
||
# distilled from Claude Sonnet 4.6 (~120k) and Claude Opus 4.6 (~10k)
|
||
# plus mixed reasoning subsets. Assistant turns ship with literal
|
||
# <think>...</think>final-answer reasoning — preserved verbatim by the
|
||
# `claude_distill` adapter so the student model learns the exact
|
||
# surface that Qwen3.6 / Qwopus3.6-27B-v1-preview emit.
|
||
#
|
||
# Used as an SFT pre-training stage to bootstrap reasoning quality
|
||
# before the elizaOS-specific training formats are layered on top.
|
||
- slug: claude-distills
|
||
repo_id: Kassadin88/Claude-Distills
|
||
normalizer: claude_distill
|
||
weight: 0.0
|
||
est_size_gb: 1.0
|
||
priority: core
|
||
license: mit
|
||
consent_basis: licensed
|
||
notes: |
|
||
`claude_distill.jsonl` is the clean unified split (~129k records,
|
||
messages = [{role:system,content:""}, {role:user,...}, {role:assistant,
|
||
"<think>...</think>final"}]). The adapter preserves <think> tags and
|
||
assigns task_type="claude_distill"; format_for_training.py uses a
|
||
minimal CLAUDE_DISTILL_SYSTEM prompt so the model learns the
|
||
thinking-mode shape from data, not a long header.
|
||
|
||
# =================== LOCAL ELIZA CORPORA =====================
|
||
# Self-hosted nubilio bot — cron-snapshot trajectories already in
|
||
# OpenAI {messages:[…]} shape. Adapter passes the assistant content
|
||
# through verbatim because the action-planner output is already
|
||
# structured elizaOS XML.
|
||
- slug: nubilio-trajectories
|
||
local_path: local-corpora/nubilio-trajectories/training-datasets
|
||
normalizer: nubilio_trajectories
|
||
weight: 3.0
|
||
est_size_gb: 0.3
|
||
priority: core
|
||
license: proprietary
|
||
consent_basis: internal_dogfood
|
||
consent_proof_uri: https://github.com/elizaOS/eliza/blob/main/packages/training/SECURITY.md#nubilio-internal-dogfood
|
||
notes: |
|
||
Cron snapshots from the self-hosted eliza bot at eliza.nubs.site
|
||
(date range 2026-04-19 → 2026-05-02, ~6.3k examples spanning
|
||
action_planner + response tasks). Drop in nubilio-trajectories/
|
||
under training/ to ingest.
|
||
|
||
# =================== NIGHTLY TRAJECTORY EXPORT (TS → Python bridge) =========
|
||
# The TS app-training plugin's nightly cron writes per-task JSONL files
|
||
# to ${ELIZA_STATE_DIR}/training/datasets/<YYYY-MM-DD>/ in canonical
|
||
# eliza_native_v1 shape, already sanitized by the runtime privacy filter.
|
||
# The `eliza_native_passthrough` adapter reads them directly — Python
|
||
# does NOT re-filter, the TS export already did. A missing or empty
|
||
# export dir emits a warning and is skipped, not treated as an error.
|
||
- slug: eliza-nightly-action-planner
|
||
source:
|
||
type: local_path
|
||
root: "${ELIZA_STATE_DIR:-~/.eliza}/training/datasets"
|
||
glob: "*/action_planner_trajectories.jsonl"
|
||
task: action_planner
|
||
normalizer: eliza_native_passthrough
|
||
weight: 1.0
|
||
est_size_gb: 0.1
|
||
priority: core
|
||
license: proprietary
|
||
consent_basis: opt_in_user_consent
|
||
consent_proof_uri: https://github.com/elizaOS/eliza/blob/main/packages/training/SECURITY.md#nightly-trajectory-opt-in
|
||
notes: |
|
||
Nightly action_planner_trajectories.jsonl exports from the TS
|
||
trajectory-export cron. One file per UTC date under
|
||
~/.eliza/training/datasets/<YYYY-MM-DD>/.
|
||
|
||
- slug: eliza-nightly-should-respond
|
||
source:
|
||
type: local_path
|
||
root: "${ELIZA_STATE_DIR:-~/.eliza}/training/datasets"
|
||
glob: "*/should_respond_trajectories.jsonl"
|
||
task: should_respond
|
||
normalizer: eliza_native_passthrough
|
||
weight: 1.0
|
||
est_size_gb: 0.05
|
||
priority: core
|
||
license: proprietary
|
||
consent_basis: opt_in_user_consent
|
||
consent_proof_uri: https://github.com/elizaOS/eliza/blob/main/packages/training/SECURITY.md#nightly-trajectory-opt-in
|
||
notes: |
|
||
Nightly should_respond_trajectories.jsonl exports from the TS
|
||
trajectory-export cron.
|
||
|
||
- slug: eliza-nightly-context-routing
|
||
source:
|
||
type: local_path
|
||
root: "${ELIZA_STATE_DIR:-~/.eliza}/training/datasets"
|
||
glob: "*/context_routing_trajectories.jsonl"
|
||
task: context_routing
|
||
normalizer: eliza_native_passthrough
|
||
weight: 1.0
|
||
est_size_gb: 0.05
|
||
priority: core
|
||
license: proprietary
|
||
consent_basis: opt_in_user_consent
|
||
consent_proof_uri: https://github.com/elizaOS/eliza/blob/main/packages/training/SECURITY.md#nightly-trajectory-opt-in
|
||
notes: |
|
||
Nightly context_routing_trajectories.jsonl exports from the TS
|
||
trajectory-export cron.
|
||
|
||
- slug: eliza-nightly-response
|
||
source:
|
||
type: local_path
|
||
root: "${ELIZA_STATE_DIR:-~/.eliza}/training/datasets"
|
||
glob: "*/response_trajectories.jsonl"
|
||
task: response
|
||
normalizer: eliza_native_passthrough
|
||
weight: 1.0
|
||
est_size_gb: 0.1
|
||
priority: core
|
||
license: proprietary
|
||
consent_basis: opt_in_user_consent
|
||
consent_proof_uri: https://github.com/elizaOS/eliza/blob/main/packages/training/SECURITY.md#nightly-trajectory-opt-in
|
||
notes: |
|
||
Nightly response_trajectories.jsonl exports from the TS
|
||
trajectory-export cron.
|
||
|
||
- slug: eliza-nightly-media-description
|
||
source:
|
||
type: local_path
|
||
root: "${ELIZA_STATE_DIR:-~/.eliza}/training/datasets"
|
||
glob: "*/media_description_trajectories.jsonl"
|
||
task: media_description
|
||
normalizer: eliza_native_passthrough
|
||
weight: 1.0
|
||
est_size_gb: 0.05
|
||
priority: core
|
||
license: proprietary
|
||
consent_basis: opt_in_user_consent
|
||
consent_proof_uri: https://github.com/elizaOS/eliza/blob/main/packages/training/SECURITY.md#nightly-trajectory-opt-in
|
||
notes: |
|
||
Nightly media_description_trajectories.jsonl exports from the TS
|
||
trajectory-export cron.
|
||
|
||
# Scam defense corpus alias. The former proprietary
|
||
# local-corpora/full-corpus-unweighted source is replaced by the public
|
||
# ScamBench training release.
|
||
- slug: scam-defense-corpus
|
||
repo_id: shaw/scambench-training
|
||
normalizer: scambench_passthrough
|
||
weight: 4.0
|
||
est_size_gb: 0.5
|
||
priority: core
|
||
license: cc-by-sa-4.0
|
||
consent_basis: licensed
|
||
notes: |
|
||
Public ScamBench training release used for the scam-defense corpus
|
||
slot. Pulls the canonical `formats/eliza-{train,test,validation}.jsonl`
|
||
files from shaw/scambench-training.
|
||
|
||
# Multi-party fantasy roleplay — Facebook LIGHT MultiLIGHT.
|
||
- slug: light-multilight
|
||
local_path: local-corpora/light-multilight/processed
|
||
normalizer: light_multilight
|
||
weight: 2.0
|
||
est_size_gb: 0.2
|
||
priority: core
|
||
license: mit
|
||
consent_basis: licensed
|
||
notes: |
|
||
Multi-party text-adventure dialogues from Facebook LIGHT
|
||
(https://github.com/facebookresearch/LIGHT). Source tarball
|
||
http://parl.ai/downloads/projects/multilight/multilight.tar.gz
|
||
contains an EpisodeDB + per-episode JSON dumps; pre-processed by
|
||
`local-corpora/light-multilight/preprocess.py` into split JSONLs
|
||
with one conversation per line ({episode_id, location, characters,
|
||
messages}). 10,917 episodes / ~324k turns / 3 named characters
|
||
per episode. Each turn yields one positive routing record
|
||
(RESPOND for the actual speaker), N-1 negative routing records
|
||
(IGNORE for the other characters at that turn), and one reply
|
||
record for the actual speaker.
|
||
|
||
# =================== TOOL CALLING (CORE) =======================
|
||
- slug: hermes-fc-v1
|
||
repo_id: NousResearch/hermes-function-calling-v1
|
||
normalizer: hermes_fc
|
||
weight: 2.0
|
||
est_size_gb: 0.3
|
||
priority: core
|
||
license: apache-2.0
|
||
consent_basis: licensed
|
||
|
||
- slug: hermes-fc-thinking-v1
|
||
repo_id: Jofthomas/hermes-function-calling-thinking-V1
|
||
normalizer: hermes_fc_thinking
|
||
weight: 2.0
|
||
est_size_gb: 0.4
|
||
priority: core
|
||
license: apache-2.0
|
||
consent_basis: licensed
|
||
|
||
- slug: glaive-fc-v2
|
||
repo_id: glaiveai/glaive-function-calling-v2
|
||
normalizer: glaive_fc
|
||
weight: 1.5
|
||
est_size_gb: 0.5
|
||
priority: core
|
||
license: apache-2.0
|
||
consent_basis: licensed
|
||
|
||
- slug: glaive-fc-v2-reasoning
|
||
repo_id: Magneto/glaive-function-calling-v2-reasoning-nemotron49b
|
||
normalizer: glaive_fc_reasoning
|
||
weight: 1.5
|
||
est_size_gb: 1.5
|
||
priority: core
|
||
license: apache-2.0
|
||
consent_basis: licensed
|
||
|
||
- slug: openhermes-2.5
|
||
repo_id: teknium/OpenHermes-2.5
|
||
normalizer: hermes_fc
|
||
weight: 1.0
|
||
est_size_gb: 4.0
|
||
priority: extra
|
||
license: apache-2.0
|
||
consent_basis: licensed
|
||
|
||
- slug: hermes-fc-v2
|
||
repo_id: NousResearch/Hermes-Function-Calling-V2
|
||
normalizer: hermes_fc
|
||
weight: 2.0
|
||
est_size_gb: 0.8
|
||
priority: core
|
||
license: apache-2.0
|
||
consent_basis: licensed
|
||
|
||
- slug: openclaw-operator
|
||
repo_id: CyberAGI/openclaw-operator-data
|
||
normalizer: openclaw
|
||
weight: 1.5
|
||
est_size_gb: 0.5
|
||
priority: core
|
||
license: apache-2.0
|
||
consent_basis: licensed
|
||
|
||
# xlam-fc-60k: gated repo (CC-BY-4.0 + APIGen citation) — requires HF token
|
||
# with access grant. Use convert_xlam_to_eliza.py with --hf-token when available.
|
||
# The xlam_fc adapter in lib/adapters.py handles the format.
|
||
|
||
- slug: sharegpt-tool-calls
|
||
repo_id: Guilherme34/sharegpt-tool-calls
|
||
normalizer: sharegpt_tool_calls
|
||
weight: 1.0
|
||
est_size_gb: 0.2
|
||
priority: core
|
||
license: unknown
|
||
consent_basis: licensed
|
||
|
||
- slug: functions-53k
|
||
repo_id: crumb/Functions-53K
|
||
normalizer: functions_53k
|
||
weight: 1.0
|
||
est_size_gb: 0.1
|
||
priority: core
|
||
license: unknown
|
||
consent_basis: licensed
|
||
|
||
- slug: bitagent-tool-calling
|
||
repo_id: BitAgent/tool_calling
|
||
normalizer: bitagent
|
||
weight: 1.0
|
||
est_size_gb: 0.3
|
||
priority: core
|
||
license: unknown
|
||
consent_basis: licensed
|
||
|
||
- slug: toolhop
|
||
repo_id: bytedance-research/ToolHop
|
||
normalizer: toolhop
|
||
weight: 1.0
|
||
est_size_gb: 0.2
|
||
priority: core
|
||
license: cc-by-4.0
|
||
consent_basis: licensed
|
||
|
||
# =================== AGENTIC / OPERATOR / MOBILE ================
|
||
- slug: mobile-actions
|
||
repo_id: google/mobile-actions
|
||
normalizer: mobile_actions
|
||
weight: 1.0
|
||
est_size_gb: 1.5
|
||
priority: core
|
||
license: cc-by-4.0
|
||
consent_basis: licensed
|
||
|
||
- slug: nemotron-rl-tool-use
|
||
repo_id: nvidia/Nemotron-RL-Agentic-Conversational-Tool-Use-Pivot-v1
|
||
normalizer: nemotron_rl_tool_use
|
||
weight: 1.5
|
||
est_size_gb: 1.0
|
||
priority: core
|
||
license: nvidia
|
||
consent_basis: licensed
|
||
|
||
- slug: qwen36-trajectory
|
||
repo_id: zake7749/Qwen-3.6-plus-agent-tool-calling-trajectory
|
||
normalizer: qwen36_trajectory
|
||
weight: 2.0
|
||
est_size_gb: 1.0
|
||
priority: core
|
||
license: unknown
|
||
consent_basis: licensed
|
||
|
||
# =================== HERMES REASONING / TRACES =================
|
||
- slug: hermes-reasoning-tool-use
|
||
repo_id: interstellarninja/hermes_reasoning_tool_use
|
||
normalizer: hermes_reasoning_tool_use
|
||
weight: 1.5
|
||
est_size_gb: 0.5
|
||
priority: core
|
||
license: apache-2.0
|
||
consent_basis: licensed
|
||
|
||
- slug: dolci-instruct-tool-use
|
||
repo_id: allenai/Dolci-Instruct-SFT-Tool-Use
|
||
normalizer: dolci_instruct
|
||
weight: 1.5
|
||
est_size_gb: 1.0
|
||
priority: core
|
||
license: odc-by
|
||
consent_basis: licensed
|
||
|
||
- slug: tool-reasoning-coding-nemotron
|
||
repo_id: AmanPriyanshu/tool-reasoning-sft-CODING-nvidia-Nemotron-Agentic-v1
|
||
normalizer: nemotron_coding_reasoning
|
||
weight: 1.0
|
||
est_size_gb: 0.7
|
||
priority: extra
|
||
license: nvidia
|
||
consent_basis: licensed
|
||
|
||
- slug: hf-coding-tools-traces
|
||
repo_id: clem/hf-coding-tools-traces
|
||
normalizer: hf_coding_tools_traces
|
||
weight: 1.0
|
||
est_size_gb: 0.5
|
||
priority: extra
|
||
license: unknown
|
||
consent_basis: licensed
|
||
|
||
- slug: hermes-agent-reasoning-traces
|
||
repo_id: lambda/hermes-agent-reasoning-traces
|
||
normalizer: hermes_traces
|
||
weight: 1.0
|
||
est_size_gb: 0.7
|
||
priority: extra
|
||
license: unknown
|
||
consent_basis: licensed
|
||
|
||
- slug: nemotron-nano-hermes-traces
|
||
repo_id: ning423/nemotron-nano-hermes-traces
|
||
normalizer: hermes_traces
|
||
weight: 1.0
|
||
est_size_gb: 0.5
|
||
priority: extra
|
||
license: unknown
|
||
consent_basis: licensed
|
||
|
||
- slug: hermes-omniforge-qwen36
|
||
repo_id: ning423/Hermes-OmniForge-Qwen36-27B-full-v0.3.0-unsloth
|
||
normalizer: hermes_omniforge
|
||
weight: 1.5
|
||
est_size_gb: 2.0
|
||
priority: extra
|
||
license: unknown
|
||
consent_basis: licensed
|
||
|
||
- slug: hermes-agent-traces-filtered
|
||
repo_id: DJLougen/hermes-agent-traces-filtered
|
||
normalizer: hermes_traces
|
||
weight: 1.0
|
||
est_size_gb: 0.5
|
||
priority: extra
|
||
license: unknown
|
||
consent_basis: licensed
|
||
|
||
- slug: aureth-corpus-hermes
|
||
repo_id: OusiaResearch/Aureth-Corpus-Hermes4.3-Generated
|
||
normalizer: aureth
|
||
weight: 1.0
|
||
est_size_gb: 1.0
|
||
priority: extra
|
||
license: unknown
|
||
consent_basis: licensed
|
||
|
||
- slug: hermes-3
|
||
repo_id: NousResearch/Hermes-3-Dataset
|
||
normalizer: hermes_3
|
||
weight: 1.0
|
||
est_size_gb: 2.0
|
||
priority: extra
|
||
license: apache-2.0
|
||
consent_basis: licensed
|
||
|
||
- slug: carnice-glm5-hermes
|
||
repo_id: kai-os/carnice-glm5-hermes-traces
|
||
normalizer: hermes_traces
|
||
weight: 1.0
|
||
est_size_gb: 0.5
|
||
priority: extra
|
||
license: unknown
|
||
consent_basis: licensed
|
||
|
||
- slug: talos-kimi-hermes
|
||
repo_id: DJLougen/Talos-kimi-k2.6-Hermes-synthetic
|
||
normalizer: hermes_traces
|
||
weight: 1.0
|
||
est_size_gb: 0.5
|
||
priority: extra
|
||
license: unknown
|
||
consent_basis: licensed
|
||
|
||
# =================== MCP TOOL CALLING ===========================
|
||
- slug: mcp-flow
|
||
repo_id: wwh0411/MCP-Flow
|
||
normalizer: mcp_flow
|
||
weight: 1.5
|
||
est_size_gb: 0.5
|
||
priority: core
|
||
license: unknown
|
||
consent_basis: licensed
|
||
notes: |
|
||
function_call/<provider>/<tool>.json files; each is a per-tool spec
|
||
with examples. Custom adapter unrolls examples into ElizaRecords.
|
||
|
||
- slug: mcp-memory-auto-trigger
|
||
repo_id: PiGrieco/mcp-memory-auto-trigger-ultimate
|
||
normalizer: mcp_messages
|
||
weight: 1.0
|
||
est_size_gb: 0.3
|
||
priority: core
|
||
license: unknown
|
||
consent_basis: licensed
|
||
|
||
- slug: mcp-agent-training-data
|
||
repo_id: muhammadtlha944/mcp-agent-training-data
|
||
normalizer: mcp_messages
|
||
weight: 1.0
|
||
est_size_gb: 0.3
|
||
priority: core
|
||
license: unknown
|
||
consent_basis: licensed
|
||
|
||
- slug: limbic-eval-tool-use-mcp
|
||
repo_id: quotientai/limbic-eval-tool-use-mcp
|
||
normalizer: mcp_messages
|
||
weight: 0.5
|
||
est_size_gb: 0.1
|
||
priority: extra
|
||
license: unknown
|
||
consent_basis: licensed
|
||
notes: eval-only, single test split
|
||
|
||
- slug: deepfabric-github-mcp
|
||
repo_id: alwaysfurther/deepfabric-github-mcp
|
||
normalizer: mcp_messages
|
||
weight: 1.0
|
||
est_size_gb: 0.5
|
||
priority: core
|
||
license: unknown
|
||
consent_basis: licensed
|
||
|
||
- slug: playwright-mcp-toolcalling
|
||
repo_id: jdaddyalbs/playwright-mcp-toolcalling
|
||
normalizer: mcp_messages
|
||
weight: 1.5
|
||
est_size_gb: 0.5
|
||
priority: core
|
||
license: unknown
|
||
consent_basis: licensed
|
||
|
||
- slug: mcp-routing-dataset
|
||
repo_id: dranganath76/mcp-routing-dataset
|
||
normalizer: mcp_routing
|
||
weight: 0.0
|
||
est_size_gb: 0.1
|
||
priority: core
|
||
license: unknown
|
||
consent_basis: licensed
|
||
|
||
- slug: ha-mcp-dataset
|
||
repo_id: jones2748/ha-mcp-dataset
|
||
normalizer: gemma_text
|
||
weight: 1.0
|
||
est_size_gb: 0.2
|
||
priority: core
|
||
license: unknown
|
||
consent_basis: licensed
|
||
|
||
- slug: phi3-mcp
|
||
repo_id: emrecandan0/phi3_mcp_dataset
|
||
normalizer: mcp_messages
|
||
weight: 1.0
|
||
est_size_gb: 0.1
|
||
priority: core
|
||
license: unknown
|
||
consent_basis: licensed
|
||
|
||
# =================== AGENT TRAJECTORIES / SHELL ====================
|
||
- slug: agent-trove
|
||
repo_id: open-thoughts/AgentTrove
|
||
normalizer: agent_trove
|
||
weight: 1.5
|
||
est_size_gb: 1.5
|
||
priority: core
|
||
license: unknown
|
||
consent_basis: licensed
|
||
|
||
- slug: nemotron-terminal-corpus
|
||
repo_id: laion/nemotron-terminal-corpus-unified
|
||
normalizer: terminal_corpus
|
||
weight: 2.0
|
||
est_size_gb: 1.5
|
||
priority: core
|
||
license: unknown
|
||
consent_basis: licensed
|
||
notes: SHELL_COMMAND training corpus.
|
||
|
||
# =================== REASONING / CoT (Jackrong) ====================
|
||
- slug: deepseek-v4-distill-8000x
|
||
repo_id: Jackrong/DeepSeek-V4-Distill-8000x
|
||
normalizer: reasoning_cot
|
||
weight: 1.0
|
||
est_size_gb: 0.5
|
||
priority: extra
|
||
license: unknown
|
||
consent_basis: licensed
|
||
|
||
- slug: glm-51-reasoning-1m
|
||
repo_id: Jackrong/GLM-5.1-Reasoning-1M-Cleaned
|
||
normalizer: reasoning_cot
|
||
weight: 0.5
|
||
est_size_gb: 3.0
|
||
priority: extra
|
||
license: unknown
|
||
consent_basis: licensed
|
||
|
||
- slug: kimi-k25-reasoning-1m
|
||
repo_id: Jackrong/Kimi-K2.5-Reasoning-1M-Cleaned
|
||
normalizer: reasoning_cot
|
||
weight: 0.5
|
||
est_size_gb: 3.0
|
||
priority: extra
|
||
license: unknown
|
||
consent_basis: licensed
|
||
|
||
- slug: qwen35-reasoning-700x
|
||
repo_id: Jackrong/Qwen3.5-reasoning-700x
|
||
normalizer: reasoning_cot
|
||
weight: 1.0
|
||
est_size_gb: 0.5
|
||
priority: extra
|
||
license: unknown
|
||
consent_basis: licensed
|
||
|
||
- slug: glm-47-multiturn-cot
|
||
repo_id: Jackrong/glm-4.7-multiturn-CoT
|
||
normalizer: reasoning_cot
|
||
weight: 1.0
|
||
est_size_gb: 1.0
|
||
priority: extra
|
||
license: unknown
|
||
consent_basis: licensed
|
||
|
||
# =================== TOOL USE EXTRAS ===============================
|
||
- slug: open-paws-tool-use
|
||
repo_id: open-paws/tool-use-llama-format
|
||
normalizer: open_paws_llama
|
||
weight: 1.0
|
||
est_size_gb: 0.5
|
||
priority: core
|
||
license: unknown
|
||
consent_basis: licensed
|
||
|
||
- slug: tool-reasoning-toucan
|
||
repo_id: AmanPriyanshu/tool-reasoning-sft-TOOLS-toucan-1.5m-sft-tool-use-data-cleaned-rectified-333k
|
||
normalizer: hermes_traces
|
||
weight: 1.5
|
||
est_size_gb: 2.0
|
||
priority: core
|
||
license: unknown
|
||
consent_basis: licensed
|
||
|
||
- slug: tool-reasoning-jupyter
|
||
repo_id: AmanPriyanshu/tool-reasoning-sft-CODING-jupyter-agent-dataset-sft-tool-use-agent-data-cleaned-rectified
|
||
normalizer: hermes_traces
|
||
weight: 1.0
|
||
est_size_gb: 0.5
|
||
priority: extra
|
||
license: unknown
|
||
consent_basis: licensed
|
||
|
||
- slug: akicou-merged-tool-use
|
||
repo_id: Akicou/merged-tool-use
|
||
normalizer: hermes_traces
|
||
weight: 1.0
|
||
est_size_gb: 0.5
|
||
priority: extra
|
||
license: unknown
|
||
consent_basis: licensed
|
||
|
||
- slug: hermes-reasoning-mlx-goekdeniz
|
||
repo_id: Goekdeniz-Guelmez/Hermes-Reasoning-Tool-Use-MLX
|
||
normalizer: hermes_traces
|
||
weight: 1.0
|
||
est_size_gb: 0.5
|
||
priority: extra
|
||
license: unknown
|
||
consent_basis: licensed
|
||
|
||
- slug: hermes-reasoning-mlx-community
|
||
repo_id: mlx-community/hermes-reasoning-tool-use
|
||
normalizer: hermes_traces
|
||
weight: 1.0
|
||
est_size_gb: 0.5
|
||
priority: extra
|
||
license: unknown
|
||
consent_basis: licensed
|
||
|
||
- slug: hermes-reasoning-cursing
|
||
repo_id: patrickleenyc/hermes_reasoning_tool_use_with_cursing
|
||
normalizer: hermes_traces
|
||
weight: 0.5
|
||
est_size_gb: 0.3
|
||
priority: extra
|
||
license: unknown
|
||
consent_basis: licensed
|
||
|
||
- slug: tool-use-multiturn-thinking
|
||
repo_id: AIGym/tool-use-multiturn-thinking
|
||
normalizer: hermes_traces
|
||
weight: 1.5
|
||
est_size_gb: 1.0
|
||
priority: core
|
||
license: unknown
|
||
consent_basis: licensed
|
||
|
||
- slug: easyr1-pc-agent-omniparser
|
||
repo_id: mlfoundations-cua-dev/easyr1-pc-agent-e-only-4k9-omniparser-qwen-tool-call-4MP
|
||
normalizer: hermes_traces
|
||
weight: 1.0
|
||
est_size_gb: 1.5
|
||
priority: extra
|
||
license: unknown
|
||
consent_basis: licensed
|
||
|
||
- slug: sage-agent-tools
|
||
repo_id: intellistream/sage-agent-tools
|
||
normalizer: hermes_traces
|
||
weight: 1.0
|
||
est_size_gb: 0.5
|
||
priority: extra
|
||
license: unknown
|
||
consent_basis: licensed
|
||
|
||
- slug: tool-call-ack-agent
|
||
repo_id: ChandraAdhikarla/tool-call-ack-dataset-agent
|
||
normalizer: hermes_traces
|
||
weight: 1.0
|
||
est_size_gb: 0.3
|
||
priority: extra
|
||
license: unknown
|
||
consent_basis: licensed
|
||
|
||
- slug: agent-conversations-retrieval
|
||
repo_id: jamescalam/agent-conversations-retrieval-tool
|
||
normalizer: hermes_traces
|
||
weight: 1.0
|
||
est_size_gb: 0.3
|
||
priority: extra
|
||
license: unknown
|
||
consent_basis: licensed
|
||
|
||
- slug: agent-tool-use-dialogue-mgmgrand
|
||
repo_id: Mgmgrand420/Agent-Tool-Use-Dialogue-Open-Dataset
|
||
normalizer: hermes_traces
|
||
weight: 1.0
|
||
est_size_gb: 0.3
|
||
priority: extra
|
||
license: unknown
|
||
consent_basis: licensed
|
||
|
||
- slug: agent-tool-use-dialogue-deepnlp
|
||
repo_id: DeepNLP/Agent-Tool-Use-Dialogue-Open-Dataset
|
||
normalizer: hermes_traces
|
||
weight: 1.0
|
||
est_size_gb: 0.3
|
||
priority: extra
|
||
license: unknown
|
||
consent_basis: licensed
|
||
|
||
# =================== RAW DIALOGUE (consumed by synthesize_routing.py)
|
||
# These do NOT yield supervised records during normalize; the routing
|
||
# synthesizer mixes them into multi-thread RESPOND/IGNORE training data.
|
||
- slug: discord-chat
|
||
repo_id: breadlicker45/discord-chat
|
||
normalizer: dialogue_raw
|
||
weight: 0.0
|
||
est_size_gb: 1.0
|
||
priority: core
|
||
license: unknown
|
||
consent_basis: licensed
|
||
notes: raw — used by synthesize_routing.py
|
||
|
||
- slug: telegram-filtered-messages
|
||
repo_id: TomData/telegram-filtered-messages
|
||
normalizer: dialogue_raw
|
||
weight: 0.0
|
||
est_size_gb: 1.0
|
||
priority: core
|
||
license: unknown
|
||
consent_basis: licensed
|
||
notes: raw — used by synthesize_routing.py
|
||
|
||
- slug: discord-dialogues
|
||
repo_id: mookiezi/Discord-Dialogues
|
||
normalizer: dialogue_raw
|
||
weight: 0.0
|
||
est_size_gb: 1.0
|
||
priority: core
|
||
license: unknown
|
||
consent_basis: licensed
|
||
notes: raw — used by synthesize_routing.py
|
||
|
||
# ============== AGENT/TOOL-USE EXTRAS (not yet in registry) ==============
|
||
- slug: pi-mono
|
||
repo_id: badlogicgames/pi-mono
|
||
normalizer: hermes_traces
|
||
weight: 1.0
|
||
est_size_gb: 0.5
|
||
priority: extra
|
||
license: unknown
|
||
consent_basis: licensed
|
||
|
||
- slug: monodox-agent-tool-use
|
||
repo_id: monodox/agent-conversations-and-tool-use
|
||
normalizer: hermes_traces
|
||
weight: 1.0
|
||
est_size_gb: 0.5
|
||
priority: extra
|
||
license: unknown
|
||
consent_basis: licensed
|
||
|
||
# ============== OPUS / FRONTIER REASONING DISTILLATIONS ==================
|
||
# Many of these are the same source data re-uploaded by different users; the
|
||
# pack-time hash dedupe collapses identical records, so weighting & priorities
|
||
# are matched to "raw signal" not "republish count".
|
||
- slug: opus-4647-reasoning-8k7
|
||
repo_id: angrygiraffe/claude-opus-4.6-4.7-reasoning-8.7k
|
||
normalizer: reasoning_cot
|
||
weight: 1.5
|
||
est_size_gb: 0.5
|
||
priority: core
|
||
license: unknown
|
||
consent_basis: licensed
|
||
|
||
- slug: swebench-verified-opus-47
|
||
repo_id: SWE-Router/swebench-verified-claude-opus-4.7
|
||
normalizer: reasoning_cot
|
||
weight: 2.0
|
||
est_size_gb: 1.5
|
||
priority: core
|
||
license: mit
|
||
consent_basis: licensed
|
||
notes: SWE-bench verified solved with Opus 4.7 — strong coding signal.
|
||
|
||
- slug: opus-47-max-sft-labs
|
||
repo_id: labs-aibusiness/reasoning-distill-opus-4-7-max-sft
|
||
normalizer: chatml_text
|
||
weight: 1.0
|
||
est_size_gb: 1.5
|
||
priority: core
|
||
license: unknown
|
||
consent_basis: licensed
|
||
|
||
- slug: opus-47-max-bas95
|
||
repo_id: Bas95/reasoning-distill-claude-opus-4-7-max
|
||
normalizer: reasoning_cot
|
||
weight: 0.5
|
||
est_size_gb: 1.5
|
||
priority: extra
|
||
license: unknown
|
||
consent_basis: licensed
|
||
|
||
- slug: opus-46-10kx-bas95
|
||
repo_id: Bas95/claude-opus-4.6-10000x
|
||
normalizer: reasoning_cot
|
||
weight: 1.0
|
||
est_size_gb: 1.5
|
||
priority: core
|
||
license: unknown
|
||
consent_basis: licensed
|
||
|
||
- slug: opus-47-reachy-app-clem
|
||
repo_id: clem/opus_4.7_buildmyfirstreachyminiapp
|
||
normalizer: reasoning_cot
|
||
weight: 1.0
|
||
est_size_gb: 0.3
|
||
priority: extra
|
||
license: unknown
|
||
consent_basis: licensed
|
||
|
||
- slug: opus-46-10kx-knowurknottty
|
||
repo_id: knowurknottty/claude-opus-4.6-10000x
|
||
normalizer: reasoning_cot
|
||
weight: 0.5
|
||
est_size_gb: 1.5
|
||
priority: extra
|
||
license: unknown
|
||
consent_basis: licensed
|
||
|
||
- slug: opus-47-reachy-mini-clem
|
||
repo_id: clem/opus4.7_reachy_mini_app_building
|
||
normalizer: reasoning_cot
|
||
weight: 1.0
|
||
est_size_gb: 0.3
|
||
priority: extra
|
||
license: unknown
|
||
consent_basis: licensed
|
||
|
||
- slug: opus-46-reasoning-12k-senaro
|
||
repo_id: senaro/opus-4.6-reasoning-sft-12k
|
||
normalizer: reasoning_cot
|
||
weight: 1.0
|
||
est_size_gb: 0.3
|
||
priority: extra
|
||
license: unknown
|
||
consent_basis: licensed
|
||
|
||
- slug: opus-47-thinking-25k-ansulev
|
||
repo_id: ansulev/Opus4.7_thinking_max_distill_god_seed_25k
|
||
normalizer: reasoning_cot
|
||
weight: 1.5
|
||
est_size_gb: 1.5
|
||
priority: core
|
||
license: unknown
|
||
consent_basis: licensed
|
||
|
||
- slug: zomi-opus-tatoeba
|
||
repo_id: zomi-language-corpora/English-Zomi-OPUS_Tatoeba_v20230412
|
||
normalizer: hermes_traces
|
||
weight: 0.2
|
||
est_size_gb: 0.1
|
||
priority: extra
|
||
license: unknown
|
||
consent_basis: licensed
|
||
notes: |
|
||
Helsinki OPUS translation corpus (English↔Zomi) — NOT Claude Opus.
|
||
Low weight; pack will mostly drop these unless we add a translation
|
||
adapter later.
|
||
|
||
- slug: opus-47-lordx64-cleaned
|
||
repo_id: TeichAI/lordx64-claude-opus-4.7-max-cleaned
|
||
normalizer: reasoning_cot
|
||
weight: 1.0
|
||
est_size_gb: 1.0
|
||
priority: extra
|
||
license: unknown
|
||
consent_basis: licensed
|
||
|
||
- slug: opus-47-huihui
|
||
repo_id: huihui-ai/opus-4-7
|
||
normalizer: reasoning_cot
|
||
weight: 1.0
|
||
est_size_gb: 1.0
|
||
priority: extra
|
||
license: unknown
|
||
consent_basis: licensed
|
||
|
||
- slug: noesis-1m-multimodel
|
||
repo_id: AMAImedia/NOESIS-1M-reasoning-router-code-math-psych-opus47-deepseek4-qwen36-gemini31-r1-gpt54
|
||
normalizer: noesis_text
|
||
weight: 2.0
|
||
est_size_gb: 5.0
|
||
priority: core
|
||
license: apache-2.0
|
||
consent_basis: licensed
|
||
notes: "1M-row mixed reasoning/router/code/math from many frontier models. Flat User/Assistant text dump."
|
||
|
||
- slug: opus-46-quyanh
|
||
repo_id: quyanh/opus-4.6
|
||
normalizer: reasoning_cot
|
||
weight: 1.0
|
||
est_size_gb: 1.0
|
||
priority: extra
|
||
license: unknown
|
||
consent_basis: licensed
|
||
|
||
- slug: opus-46-ptbr-3kx
|
||
repo_id: traderpedroso/PTBR-Opus-4.6-Reasoning-3000x-filtered
|
||
normalizer: reasoning_cot
|
||
weight: 0.5
|
||
est_size_gb: 0.5
|
||
priority: extra
|
||
license: unknown
|
||
consent_basis: licensed
|
||
notes: Portuguese-BR variant.
|
||
|
||
- slug: opus-47-max-sft-snow257
|
||
repo_id: Snow257/reasoning-distill-opus-4-7-max-sft
|
||
normalizer: reasoning_cot
|
||
weight: 0.3
|
||
est_size_gb: 1.0
|
||
priority: extra
|
||
license: unknown
|
||
consent_basis: licensed
|
||
|
||
- slug: loomstack-opus-distill-prompted
|
||
repo_id: Itachi-42/loomstack-opus-distill-prompted
|
||
normalizer: reasoning_cot
|
||
weight: 1.0
|
||
est_size_gb: 0.5
|
||
priority: extra
|
||
license: unknown
|
||
consent_basis: licensed
|
||
|
||
- slug: opus-46-10kx-skipharlow
|
||
repo_id: skipharlow/claude-opus-4.6-10000x
|
||
normalizer: reasoning_cot
|
||
weight: 0.3
|
||
est_size_gb: 1.5
|
||
priority: extra
|
||
license: unknown
|
||
consent_basis: licensed
|
||
|
||
- slug: loomstack-opus-distill
|
||
repo_id: mohit-1710/loomstack-opus-distill
|
||
normalizer: reasoning_cot
|
||
weight: 1.0
|
||
est_size_gb: 0.5
|
||
priority: extra
|
||
license: unknown
|
||
consent_basis: licensed
|
||
|
||
- slug: opus-46-10kx-edawking
|
||
repo_id: Edawking/claude-opus-4.6-10000x
|
||
normalizer: reasoning_cot
|
||
weight: 0.3
|
||
est_size_gb: 1.5
|
||
priority: extra
|
||
license: unknown
|
||
consent_basis: licensed
|
||
|
||
- slug: opus-47-reasoning-cot-ansulev
|
||
repo_id: ansulev/opus-4.7-reasoning-cot
|
||
normalizer: reasoning_cot
|
||
weight: 1.0
|
||
est_size_gb: 1.0
|
||
priority: core
|
||
license: unknown
|
||
consent_basis: licensed
|
||
|
||
- slug: opus-47-max-sft-ansulev
|
||
repo_id: ansulev/reasoning-distill-opus-4-7-max-sft
|
||
normalizer: reasoning_cot
|
||
weight: 0.3
|
||
est_size_gb: 1.5
|
||
priority: extra
|
||
license: unknown
|
||
consent_basis: licensed
|
||
|
||
- slug: opus-46-3kx-youssef
|
||
repo_id: Youssef001122/Opus-4.6-Reasoning-3000x-filtered
|
||
normalizer: reasoning_cot
|
||
weight: 0.3
|
||
est_size_gb: 0.5
|
||
priority: extra
|
||
license: unknown
|
||
consent_basis: licensed
|
||
|
||
- slug: opus-reasoning-filtered-octomed
|
||
repo_id: OctoMed/Opus-Reasoning-Filtered
|
||
normalizer: reasoning_cot
|
||
weight: 1.0
|
||
est_size_gb: 0.5
|
||
priority: extra
|
||
license: unknown
|
||
consent_basis: licensed
|
||
|
||
- slug: opus-47-max-sft-trilogix
|
||
repo_id: Trilogix1/reasoning-distill-opus-4-7-max-sft
|
||
normalizer: reasoning_cot
|
||
weight: 0.3
|
||
est_size_gb: 1.5
|
||
priority: extra
|
||
license: unknown
|
||
consent_basis: licensed
|
||
|
||
- slug: opus-47-thinking-25k-trilogix
|
||
repo_id: Trilogix1/Opus4.7_thinking_max_distill_god_seed_25k
|
||
normalizer: reasoning_cot
|
||
weight: 0.3
|
||
est_size_gb: 1.5
|
||
priority: extra
|
||
license: unknown
|
||
consent_basis: licensed
|
||
|
||
- slug: opus-47-thinking-25k-withinus
|
||
repo_id: WithinUsAI/Opus4.7_thinking_max_distill_god_seed_25k
|
||
normalizer: reasoning_cot
|
||
weight: 0.3
|
||
est_size_gb: 1.5
|
||
priority: extra
|
||
license: unknown
|
||
consent_basis: licensed
|
||
|
||
- slug: opus-46-10kx-jeremydh
|
||
repo_id: Jeremydh911/claude-opus-4.6-10000x
|
||
normalizer: reasoning_cot
|
||
weight: 0.3
|
||
est_size_gb: 1.5
|
||
priority: extra
|
||
license: unknown
|
||
consent_basis: licensed
|
||
|
||
- slug: opus-translation-en-de-cs
|
||
repo_id: davidmeikle/opus-translation-train-en-de-cs
|
||
normalizer: hermes_traces
|
||
weight: 0.2
|
||
est_size_gb: 0.5
|
||
priority: extra
|
||
license: unknown
|
||
consent_basis: licensed
|
||
notes: |
|
||
Helsinki OPUS translation (en→de/cs) — NOT Claude Opus. Low weight.
|
||
|
||
- slug: opus-46-3kx-bakrianto
|
||
repo_id: bakrianto/Opus-4.6-Reasoning-3000x-filtered
|
||
normalizer: reasoning_cot
|
||
weight: 0.3
|
||
est_size_gb: 0.5
|
||
priority: extra
|
||
license: unknown
|
||
consent_basis: licensed
|
||
|
||
- slug: opus-46-2kx-compactai
|
||
repo_id: CompactAI-O/Opus-4.6-Reasoning-2160x
|
||
normalizer: reasoning_cot
|
||
weight: 1.0
|
||
est_size_gb: 0.5
|
||
priority: extra
|
||
license: unknown
|
||
consent_basis: licensed
|
||
|
||
# =================== MULTI-PARTY → should_respond ================
|
||
# These corpora are NOT consumed by normalize.py. They are read by
|
||
# scripts/synthesize_multiparty_routing.py, which chunks each
|
||
# conversation, picks one participant as the "agent", and labels
|
||
# RESPOND vs IGNORE based on whether the agent actually spoke next
|
||
# in the original transcript or was directly addressed. Output goes
|
||
# to data/synthesized/multiparty_should_respond.jsonl with
|
||
# task_type="should_respond" — same shape as the runtime handler
|
||
# consumes. Weight=0 here so pack_dataset.py ignores raw rows.
|
||
|
||
- slug: ishiki-labs-multi-party-dialogue
|
||
repo_id: ishiki-labs/multi-party-dialogue
|
||
normalizer: dialogue_raw
|
||
weight: 0.0
|
||
est_size_gb: 0.4
|
||
priority: core
|
||
license: apache-2.0
|
||
consent_basis: licensed
|
||
notes: |
|
||
Pre-labeled SPEAK/SILENT decision points across AMI meetings,
|
||
Friends scripts, and SPGI calls. ~50/50 split out of the box.
|
||
Synthesizer maps SPEAK→RESPOND, SILENT→IGNORE.
|
||
|
||
- slug: mrfakename-multilight-sharegpt
|
||
repo_id: mrfakename/multilight-sharegpt
|
||
normalizer: dialogue_raw
|
||
weight: 0.0
|
||
est_size_gb: 0.05
|
||
priority: core
|
||
license: apache-2.0
|
||
consent_basis: licensed
|
||
notes: |
|
||
Multi-character ShareGPT roleplay. Synthesizer chunks each
|
||
conversation and picks a real character as the agent.
|
||
|
||
- slug: cfettuccini-multipartyconv
|
||
repo_id: CFettuccini/multipartyconv
|
||
normalizer: dialogue_raw
|
||
weight: 0.0
|
||
est_size_gb: 0.05
|
||
priority: core
|
||
license: unknown
|
||
consent_basis: licensed
|
||
notes: |
|
||
CSV of `<s>[INST]>>{user1} : msg>>{user2} : msg...` strings;
|
||
synthesizer parses the >>{name} : pattern.
|
||
|
||
- slug: nu-dialogue-multi-relational-mp
|
||
repo_id: nu-dialogue/multi-relational-multi-party-chat-corpus
|
||
normalizer: dialogue_raw
|
||
weight: 0.0
|
||
est_size_gb: 0.05
|
||
priority: core
|
||
license: cc-by-nd-4.0
|
||
consent_basis: licensed
|
||
notes: |
|
||
Japanese multi-party dialogues. The HF entry is a script-loader
|
||
that snapshot_download skips; synthesizer fetches the upstream
|
||
GitHub release zip directly on first run.
|
||
|
||
# =================== n8n WORKFLOW GENERATION =====================
|
||
# Universal n8n adapter feature-detects across six row shapes
|
||
# (messages, prompt+json, alpaca, raw workflow JSON, kv pairs, DPO).
|
||
# All emit task_type="n8n_workflow_generation". Many of these are
|
||
# republishes of the same underlying corpora — pack-time hash dedupe
|
||
# collapses identical records, so weighting is matched to raw signal.
|
||
|
||
- slug: n8n-mega-workflows
|
||
repo_id: Ker102/n8n-mega-workflows
|
||
normalizer: n8n_workflow
|
||
weight: 2.0
|
||
est_size_gb: 0.5
|
||
priority: core
|
||
license: unknown
|
||
consent_basis: licensed
|
||
notes: 131k alpaca-style {instruction,input,output} with category + complexity.
|
||
|
||
- slug: n8n-toolkit-davidrpatton
|
||
repo_id: DavidrPatton/n8n-Toolkit
|
||
normalizer: n8n_workflow
|
||
weight: 1.5
|
||
est_size_gb: 0.6
|
||
priority: core
|
||
license: unknown
|
||
consent_basis: licensed
|
||
notes: 56k messages-shape with scraped readme + screenshot metadata.
|
||
|
||
- slug: n8n-workflows-templates-0xarchit
|
||
repo_id: 0xarchit/n8n_workflows_templates_dataset
|
||
normalizer: n8n_workflow
|
||
weight: 1.0
|
||
est_size_gb: 0.8
|
||
priority: core
|
||
license: unknown
|
||
consent_basis: licensed
|
||
notes: 49k {key,value} parquet — synthetic prompt → workflow JSON.
|
||
|
||
- slug: n8n-master-corpus
|
||
repo_id: Ker102/n8n-master-corpus
|
||
normalizer: n8n_workflow
|
||
weight: 1.5
|
||
est_size_gb: 0.1
|
||
priority: core
|
||
license: unknown
|
||
consent_basis: licensed
|
||
notes: 36k raw workflow_json + integrations/category. Synthetic prompt.
|
||
|
||
- slug: n8n-workflows-sft-eclaude
|
||
repo_id: eclaude/n8n-workflows-sft
|
||
normalizer: n8n_workflow
|
||
weight: 1.5
|
||
est_size_gb: 0.5
|
||
priority: core
|
||
license: unknown
|
||
consent_basis: licensed
|
||
notes: messages-shape with train/eval splits.
|
||
|
||
- slug: n8n-workflow-dataset-ruh-ai
|
||
repo_id: ruh-ai/n8n-workflow-dataset
|
||
normalizer: n8n_workflow
|
||
weight: 1.5
|
||
est_size_gb: 0.2
|
||
priority: core
|
||
license: unknown
|
||
consent_basis: licensed
|
||
notes: prompt+json+thinking — thinking stashed in metadata.
|
||
|
||
- slug: n8n-workflows-v2-4k-arkelai
|
||
repo_id: arkelai/n8n-workflows-v2-4k
|
||
normalizer: n8n_workflow
|
||
weight: 1.5
|
||
est_size_gb: 0.1
|
||
priority: core
|
||
license: unknown
|
||
consent_basis: licensed
|
||
notes: prompt+answer pairs.
|
||
|
||
- slug: n8n-workflows-yagnik
|
||
repo_id: Yagnik-patel-12/N8n_workflows_dataset
|
||
normalizer: n8n_workflow
|
||
weight: 0.5
|
||
est_size_gb: 0.1
|
||
priority: extra
|
||
license: unknown
|
||
consent_basis: licensed
|
||
|
||
- slug: n8n-grpo-4k-aks729
|
||
repo_id: AKS729/grpo_n8n_dataset_4k
|
||
normalizer: n8n_workflow
|
||
weight: 1.0
|
||
est_size_gb: 0.1
|
||
priority: core
|
||
license: unknown
|
||
consent_basis: licensed
|
||
notes: same prompt+json+thinking shape as ruh-ai filtered set; dedupe handles overlap.
|
||
|
||
- slug: n8n-workflow-template-rubenz
|
||
repo_id: rubenz-org/n8n-workflow-template
|
||
normalizer: n8n_workflow
|
||
weight: 1.0
|
||
est_size_gb: 0.1
|
||
priority: extra
|
||
license: unknown
|
||
consent_basis: licensed
|
||
|
||
- slug: n8n-workflow-npv2k1
|
||
repo_id: npv2k1/n8n-workflow
|
||
normalizer: n8n_workflow
|
||
weight: 0.5
|
||
est_size_gb: 0.1
|
||
priority: extra
|
||
license: unknown
|
||
consent_basis: licensed
|
||
notes: small input+output pair set; high overlap with mzw2004/Di12/tahakk/fmd053131.
|
||
|
||
- slug: n8n-workflow-yonibabi
|
||
repo_id: yonibabi/n8n-workflow-dataset
|
||
normalizer: n8n_workflow
|
||
weight: 0.3
|
||
est_size_gb: 0.1
|
||
priority: extra
|
||
license: unknown
|
||
consent_basis: licensed
|
||
|
||
- slug: n8n-workflow-di12
|
||
repo_id: Di12/n8n_workflow
|
||
normalizer: n8n_workflow
|
||
weight: 0.3
|
||
est_size_gb: 0.1
|
||
priority: extra
|
||
license: unknown
|
||
consent_basis: licensed
|
||
|
||
- slug: n8n-workflow-mzw2004
|
||
repo_id: mzw2004/n8n-workflow
|
||
normalizer: n8n_workflow
|
||
weight: 0.3
|
||
est_size_gb: 0.1
|
||
priority: extra
|
||
license: unknown
|
||
consent_basis: licensed
|
||
|
||
- slug: n8n-workflow-tahakk
|
||
repo_id: tahakk/n8n-workflow
|
||
normalizer: n8n_workflow
|
||
weight: 0.3
|
||
est_size_gb: 0.1
|
||
priority: extra
|
||
license: unknown
|
||
consent_basis: licensed
|
||
|
||
- slug: n8n-workflow-fmd053131
|
||
repo_id: fmd053131/n8n-workflow
|
||
normalizer: n8n_workflow
|
||
weight: 0.3
|
||
est_size_gb: 0.1
|
||
priority: extra
|
||
license: unknown
|
||
consent_basis: licensed
|
||
|
||
- slug: n8n-workflows-batuhanilgarr
|
||
repo_id: batuhanilgarr/n8n-workflows-dataset
|
||
normalizer: n8n_workflow
|
||
weight: 1.0
|
||
est_size_gb: 0.1
|
||
priority: extra
|
||
license: unknown
|
||
consent_basis: licensed
|
||
notes: 525 raw workflows with name/nodes/connections.
|
||
|
||
- slug: n8n-grpo-2k-aks729
|
||
repo_id: AKS729/grpo_n8n_dataset_2k
|
||
normalizer: n8n_workflow
|
||
weight: 0.5
|
||
est_size_gb: 0.1
|
||
priority: extra
|
||
license: unknown
|
||
consent_basis: licensed
|
||
|
||
- slug: n8n-preped-heriman
|
||
repo_id: Heriman/n8n_preped
|
||
normalizer: n8n_workflow
|
||
weight: 0.0
|
||
est_size_gb: 0.1
|
||
priority: extra
|
||
license: unknown
|
||
consent_basis: licensed
|
||
|
||
- slug: n8n-workflow-ruh-ai
|
||
repo_id: ruh-ai/n8n-workflow
|
||
normalizer: n8n_workflow
|
||
weight: 1.0
|
||
est_size_gb: 0.1
|
||
priority: extra
|
||
license: unknown
|
||
consent_basis: licensed
|
||
notes: 2.1k prompt+json pairs.
|
||
|
||
- slug: n8n-training-mustaphal
|
||
repo_id: MustaphaL/training_data_n8n_workflows
|
||
normalizer: n8n_workflow
|
||
weight: 1.0
|
||
est_size_gb: 0.1
|
||
priority: extra
|
||
license: unknown
|
||
consent_basis: licensed
|
||
notes: ships alpaca, simple {prompt,completion}, and openai messages variants.
|
||
|
||
- slug: n8n-testset-ruh-ai
|
||
repo_id: ruh-ai/n8n-testset
|
||
normalizer: n8n_workflow
|
||
weight: 0.3
|
||
est_size_gb: 0.1
|
||
priority: extra
|
||
license: unknown
|
||
consent_basis: licensed
|
||
notes: small eval-only split.
|
||
|
||
- slug: n8n-workflows-thinking-stmasson
|
||
repo_id: stmasson/n8n-workflows-thinking
|
||
normalizer: n8n_workflow
|
||
weight: 1.5
|
||
est_size_gb: 0.5
|
||
priority: core
|
||
license: unknown
|
||
consent_basis: licensed
|
||
notes: |
|
||
Multi-format SFT/DPO/ORPO/GRPO splits with explicit <thinking> traces.
|
||
Adapter takes the SFT 'messages' rows directly and the DPO 'chosen'
|
||
branch; rejected branch is dropped.
|
||
|
||
- slug: n8nbuilder-perspicacious
|
||
repo_id: Perspicacious/n8nbuilder-n8n-workflows-dataset
|
||
normalizer: n8n_workflow
|
||
weight: 1.0
|
||
est_size_gb: 0.1
|
||
priority: extra
|
||
license: unknown
|
||
consent_basis: licensed
|
||
notes: ships both alpaca train.jsonl and openai train_openai.jsonl.
|
||
|
||
- slug: n8nbuilder-velixar
|
||
repo_id: velixar-ai/n8nbuilder-n8n-workflows-dataset
|
||
normalizer: n8n_workflow
|
||
weight: 0.5
|
||
est_size_gb: 0.1
|
||
priority: extra
|
||
license: unknown
|
||
consent_basis: licensed
|
||
notes: republish of n8nbuilder corpus; dedupe collapses overlap.
|
||
|
||
- slug: n8nbuilder-webman
|
||
repo_id: webman/n8nbuilder-n8n-workflows-dataset
|
||
normalizer: n8n_workflow
|
||
weight: 0.5
|
||
est_size_gb: 0.1
|
||
priority: extra
|
||
license: unknown
|
||
consent_basis: licensed
|
||
notes: republish of n8nbuilder corpus; dedupe collapses overlap.
|
||
|
||
# =================== ABLITERATION CALIBRATION (NOT IN TRAIN MIX) ===================
|
||
# mlabonne/harmful_behaviors + mlabonne/harmless_alpaca are calibration
|
||
# corpora for orthogonal-projection abliteration (refusal-direction
|
||
# ablation) and for the heretic eval gate. They MUST NOT enter the
|
||
# supervised train mix — weight=0.0 makes pack_dataset.py route the
|
||
# normalized records to data/abliteration/{harmful,harmless}.jsonl
|
||
# instead of train/val/test. Consumers:
|
||
# - scripts/quantization/abliteration_apply.py (--harmful-jsonl /
|
||
# --harmless-jsonl read currentMessage.content)
|
||
# - scripts/training/abliterate.py (loads via HF directly; this entry
|
||
# just keeps the corpus discoverable + license-traceable)
|
||
- slug: harmful_behaviors
|
||
repo_id: mlabonne/harmful_behaviors
|
||
normalizer: harmful_behaviors
|
||
weight: 0.0
|
||
est_size_gb: 0.01
|
||
priority: extra
|
||
license: mit
|
||
consent_basis: licensed
|
||
notes: |
|
||
Canonical harmful-prompt corpus referenced by
|
||
scripts/quantization/abliteration_apply.py (line ~68 comment).
|
||
~520 refusal-eliciting prompts. Calibration only — weight=0.0 keeps
|
||
these records out of train.jsonl. pack_dataset.py routes records
|
||
with task_type="abliteration_harmful" to
|
||
data/abliteration/harmful.jsonl.
|
||
|
||
- slug: harmless_alpaca
|
||
repo_id: mlabonne/harmless_alpaca
|
||
normalizer: harmless_alpaca
|
||
weight: 0.0
|
||
est_size_gb: 0.05
|
||
priority: extra
|
||
license: mit
|
||
consent_basis: licensed
|
||
notes: |
|
||
Paired benign-instruction corpus for orthogonal-projection
|
||
abliteration. ~1.4k benign prompts (Alpaca-derived). Calibration
|
||
only — weight=0.0 keeps these records out of train.jsonl.
|
||
pack_dataset.py routes records with task_type="abliteration_harmless"
|
||
to data/abliteration/harmless.jsonl.
|
||
|
||
# Synthesized examples generated from elizaOS prompt templates
|
||
# (handled by synthesize_targets.py, not downloaded).
|
||
synthesized:
|
||
# Scam-refusal `reply` records derived from lalalune/scambench-training
|
||
# (English subset + balanced legitimate slice). Generated by
|
||
# scripts/sources/scambench_adapter.py. Used to re-instill refusal /
|
||
# de-escalation behavior after weight-level abliteration.
|
||
#
|
||
# File: data/synthesized/scambench/scambench.jsonl
|
||
# Pack pipeline slug: synth:scambench-scambench
|
||
- task_id: scambench-scambench
|
||
n_examples: 11653
|
||
weight: 3.0
|
||
notes: |
|
||
Per-record canonical `reply` task: JSON envelope
|
||
{thought, reply:{text, channel}}. metadata.scam_category maps the
|
||
upstream scenario_category onto the canonical taxonomy (phishing,
|
||
impersonation, crypto-scam, sextortion, romance, tech-support,
|
||
wire-fraud, other) plus `none` for legitimate control records.
|
||
System prompt is the canonical `reply` template from
|
||
data/prompts/registry.json. Higher weight (3.0) so refusal
|
||
behaviour is well-represented in the SFT pass.
|
||
|
||
# Per-action message_handler coverage. Generated by
|
||
# scripts/synthesize_action_planner.py from
|
||
# /tmp/action_inventory_enriched.json — one entry per (action,
|
||
# conversation context, prompt phrasing) triple, ~30 records per
|
||
# action across all 312 actions in the eliza tree. The supervised
|
||
# target is a JSON document matching the message_handler schema
|
||
# (thought, tool_calls[]{name,params}, providers, text, simple).
|
||
- task_id: action_planner_coverage
|
||
n_examples: 9305
|
||
weight: 2.0
|
||
notes: |
|
||
Covers every Action declared across eliza/packages/agent +
|
||
eliza/packages/core + eliza/plugins. Memory-window distribution
|
||
is roughly uniform across {0, 1, 3, 8, 15} turns. ~24% of
|
||
records chain a REPLY after the target action. All records
|
||
are valid JSON structures.
|
||
|
||
- task_id: should_respond
|
||
template: should_respond.txt
|
||
n_examples: 2000
|
||
weight: 3.0
|
||
- task_id: should_respond_with_context
|
||
template: should_respond_with_context.txt
|
||
n_examples: 1500
|
||
weight: 2.0
|
||
- task_id: reply
|
||
template: reply.txt
|
||
n_examples: 2000
|
||
weight: 2.0
|
||
- task_id: should_mute_room
|
||
template: should_mute_room.txt
|
||
n_examples: 800
|
||
weight: 1.0
|
||
- task_id: should_unmute_room
|
||
template: should_unmute_room.txt
|
||
n_examples: 600
|
||
weight: 1.0
|
||
- task_id: should_follow_room
|
||
template: should_follow_room.txt
|
||
n_examples: 600
|
||
weight: 1.0
|
||
- task_id: should_unfollow_room
|
||
template: should_unfollow_room.txt
|
||
n_examples: 600
|
||
weight: 1.0
|
||
- task_id: choose_option
|
||
template: choose_option.txt
|
||
n_examples: 1000
|
||
weight: 1.0
|
||
- task_id: option_extraction
|
||
template: option_extraction.txt
|
||
n_examples: 800
|
||
weight: 1.0
|
||
- task_id: reflection
|
||
template: reflection.txt
|
||
n_examples: 800
|
||
weight: 1.0
|
||
- task_id: reflection_evaluator
|
||
template: reflection_evaluator.txt
|
||
n_examples: 600
|
||
weight: 1.0
|
||
- task_id: think
|
||
template: think.txt
|
||
n_examples: 1000
|
||
weight: 1.0
|
||
- task_id: post_action_decision
|
||
template: post_action_decision.txt
|
||
n_examples: 800
|
||
weight: 1.0
|
||
- task_id: multi_step_decision
|
||
template: multi_step_decision.txt
|
||
n_examples: 800
|
||
weight: 1.0
|
||
- task_id: extract_secrets
|
||
template: extract_secrets.txt
|
||
n_examples: 600
|
||
weight: 1.0
|
||
- task_id: update_settings
|
||
template: update_settings.txt
|
||
n_examples: 400
|
||
weight: 0.5
|
||
- task_id: update_summarization
|
||
template: update_summarization.txt
|
||
n_examples: 400
|
||
weight: 0.5
|
||
- task_id: message_classifier
|
||
template: message_classifier.txt
|
||
n_examples: 1000
|
||
weight: 1.0
|
||
- task_id: long_term_extraction
|
||
template: long_term_extraction.txt
|
||
n_examples: 200
|
||
weight: 0.5
|
||
|
||
# Chat-augmented routing records (pre-generated by
|
||
# scripts/synthesize_routing_v2.py). The packer matches the file stem
|
||
# `dialogue_routing_v2` to this entry's task_id for weighting; the
|
||
# template/n_examples fields are nominal — synthesize_targets.py does not
|
||
# generate these, the routing-v2 synthesizer does.
|
||
- task_id: dialogue_routing_v2
|
||
template: should_respond_with_context.txt
|
||
n_examples: 20000
|
||
weight: 2.0
|
||
notes: |
|
||
Name-injection / ping / username-match should_respond records mined
|
||
from discord-chat, discord-dialogues, telegram-filtered-messages.
|
||
Generated by scripts/synthesize_routing_v2.py with seed=0xE71A05.
|
||
Action mix: RESPOND ~36%, IGNORE ~56%, STOP ~8%. Signal kinds:
|
||
direct_mention, username_match, role_address, stop, different_addressee,
|
||
group_chitchat, self, quoted_reference.
|
||
|
||
# Rule-based action-pair synthesis targeting registry-v2 + actions-catalog
|
||
# (handled by `scripts/synthesize_action_pairs.py`). One jsonl per group
|
||
# under `data/synthesized/action_pairs/`. Uses canned scenario pools and
|
||
# paraphrase variants — no LLM teacher. Every record carries a JSON
|
||
# expectedResponse keyed off the prompt's `expected_keys` so the tasks
|
||
# register in PROMPT_COVERAGE.md.
|
||
- task_id: action-pairs-core-prompts
|
||
n_examples: 3800
|
||
weight: 2.5
|
||
notes: |
|
||
100 examples per core prompt template (38 templates from
|
||
registry-v2.json). File: data/synthesized/action_pairs/core-prompts.jsonl.
|
||
- task_id: action-pairs-inline-actions
|
||
n_examples: 600
|
||
weight: 1.5
|
||
notes: |
|
||
100 examples per inline-action prompt (clipboard extract + dataset-generator
|
||
should_respond, 6 templates). File:
|
||
data/synthesized/action_pairs/inline-actions.jsonl.
|
||
- task_id: action-pairs-plugin-prompts
|
||
n_examples: 2800
|
||
weight: 2.0
|
||
notes: |
|
||
100 examples per plugin prompt (28 templates: discord, evm, solana,
|
||
shell). File: data/synthesized/action_pairs/plugin-prompts.jsonl.
|
||
- task_id: action-pairs-lifeops
|
||
n_examples: 20500
|
||
weight: 1.5
|
||
notes: |
|
||
50 paraphrased variants per lifeops scenario (410 task ids across 18
|
||
scenarios). File: data/synthesized/action_pairs/lifeops.jsonl.
|
||
- task_id: action-pairs-actions-catalog
|
||
n_examples: 5920
|
||
weight: 2.0
|
||
notes: |
|
||
100 examples per structured-param action (37) + 30 per no-param action
|
||
(74). 110 unique action names covered. File:
|
||
data/synthesized/action_pairs/actions-catalog.jsonl.
|
||
|
||
# Music plugin tool-call examples generated by
|
||
# scripts/synthesize_music_actions.py. Targets the 15 actions across
|
||
# plugin-music-library (8) + plugin-music-player (7). 100 records per
|
||
# action (1,500 total) with ~73% English / ~27% multilingual
|
||
# (zh/es/fr/ja/de/pt) and ~5% subtle-null records that resolve to a
|
||
# canonical thought:/text: REPLY shape instead of a tool call.
|
||
- task_id: action_examples-music
|
||
n_examples: 1500
|
||
weight: 2.0
|
||
notes: |
|
||
100 records per music action (15 actions, 1,500 total). File:
|
||
data/synthesized/action_examples/music.jsonl. Multilingual coverage
|
||
across zh/es/fr/ja/de/pt; ~5% subtle-null REPLY records. The slug
|
||
matches the pack-time `<parent_dir>-<stem>` derivation in
|
||
scripts/pack_dataset.py so the weight applies during sampling.
|
||
|
||
# Agent / orchestration tool-call examples generated by
|
||
# scripts/synthesize_agent_orch_actions.py. Covers 24 actions across the
|
||
# agent/orchestration plugin surface — plugin-agent-orchestrator (10),
|
||
# plugin-agent-skills (7), plugin-app-control (1, polymorphic APP),
|
||
# plugin-plugin-manager (2, polymorphic PLUGIN),
|
||
# plugin-elizacloud (4). 100 records per action (2,400 total).
|
||
# Polymorphic actions (APP, PLUGIN) spread examples across all
|
||
# documented sub-operations. ~70% English / ~30% multilingual
|
||
# (zh/es/fr/ja/de/pt) with >=3 records per non-en lang per action.
|
||
# 7% subtle-null records resolve to a canonical thought:/text: REPLY
|
||
# shape instead of a tool call. 12+ user phrasing styles.
|
||
- task_id: action_examples-agent_orch
|
||
n_examples: 2400
|
||
weight: 2.0
|
||
notes: |
|
||
100 records per action (24 actions, 2,400 total). File:
|
||
data/synthesized/action_examples/agent_orch.jsonl. Covers
|
||
plugin-agent-orchestrator, plugin-agent-skills, plugin-app-control,
|
||
plugin-plugin-manager, plugin-elizacloud. Multilingual coverage
|
||
across zh/es/fr/ja/de/pt; ~7% subtle-null REPLY records. The slug
|
||
matches the pack-time `<parent_dir>-<stem>` derivation in
|
||
scripts/pack_dataset.py so the weight applies during sampling.
|
||
|
||
# System / computer-use / github tool-call examples generated by
|
||
# scripts/synthesize_system_actions.py. Targets 11 actions across three
|
||
# plugins:
|
||
# plugin-shell (1) CLEAR_SHELL_HISTORY
|
||
# plugin-computeruse (5) BROWSER_ACTION, FILE_ACTION, MANAGE_WINDOW,
|
||
# TERMINAL_ACTION, USE_COMPUTER
|
||
# plugin-github (5) ASSIGN_ISSUE, CREATE_ISSUE,
|
||
# GITHUB_NOTIFICATION_TRIAGE, LIST_PRS,
|
||
# REVIEW_PR
|
||
# 100 records per action (1,100 total) with 70% English / 30%
|
||
# multilingual (zh/es/fr/ja/de/pt — exactly 5 per language per action)
|
||
# and ~7% subtle-null REPLY records that emit a thought:/text: shape
|
||
# when the user message is too underspecified to resolve to a tool
|
||
# call. Polymorphic *_ACTION / USE_COMPUTER actions cover every
|
||
# documented sub-operation: BROWSER_ACTION (open/navigate/click/type/
|
||
# scroll/screenshot/dom/clickables/execute/state/wait/list_tabs/
|
||
# open_tab/close_tab/switch_tab/close), FILE_ACTION (read/write/
|
||
# append/edit/delete/exists/list/delete_directory/download/upload/
|
||
# list_downloads), MANAGE_WINDOW (list/focus/switch/arrange/move/
|
||
# minimize/maximize/restore/close), TERMINAL_ACTION (connect/execute/
|
||
# execute_command/read/type/clear/close), USE_COMPUTER (screenshot/
|
||
# click/click_with_modifiers/double_click/right_click/mouse_move/
|
||
# type/key/key_combo/scroll/drag).
|
||
- task_id: action_examples-system
|
||
n_examples: 1100
|
||
weight: 2.0
|
||
notes: |
|
||
100 records per system / computer-use / github action (11 actions,
|
||
1,100 total). File: data/synthesized/action_examples/system.jsonl.
|
||
Multilingual coverage zh/es/fr/ja/de/pt (5 per language per
|
||
action); ~7% subtle-null REPLY records. Slug matches the pack-time
|
||
`<parent_dir>-<stem>` derivation in scripts/pack_dataset.py so the
|
||
weight applies during sampling.
|
||
|
||
# Messaging-plugin tool-call examples generated by
|
||
# scripts/synthesize_messaging_actions.py. Targets 34 connector intents
|
||
# across six connectors, each emitted with canonical action names; several
|
||
# intents share the same canonical action.
|
||
# plugin-discord (17 intents) CHAT_WITH_ATTACHMENTS, CREATE_POLL,
|
||
# DELETE_MESSAGE, DOWNLOAD_MEDIA, EDIT_MESSAGE,
|
||
# GET_USER, JOIN_CHANNEL, LEAVE_CHANNEL,
|
||
# LIST_CHANNELS, PIN_MESSAGE, MESSAGE,
|
||
# SERVER_INFO, SETUP_CREDENTIALS,
|
||
# SUMMARIZE_CONVERSATION, TRANSCRIBE_MEDIA
|
||
# plugin-twitter (7 intents) READ_FEED, POST, READ_MESSAGES,
|
||
# MESSAGE, SEARCH_POSTS
|
||
# plugin-signal (5 intents) GET_USER, LIST_CHANNELS,
|
||
# READ_MESSAGES,
|
||
# MESSAGE, REACT_TO_MESSAGE
|
||
# plugin-bluebubbles (2 intents) REACT_TO_MESSAGE,
|
||
# MESSAGE
|
||
# plugin-imessage (1 intent) MESSAGE
|
||
# plugin-whatsapp (2 intents) MESSAGE, REACT_TO_MESSAGE
|
||
# 100 records per connector intent (3,400 total). 70% English /
|
||
# 30% multilingual (zh/es/fr/ja/de/pt — exactly 5 per language per
|
||
# intent) across 10
|
||
# phrasing styles (direct, formal, casual, expert, naive, voice-asr,
|
||
# distracted, broken-english, self-correcting, subtle-null). 10% of
|
||
# records per intent are subtle-null cases that emit a canonical
|
||
# thought:/text: REPLY or thought:/tool_calls[0] IGNORE shape rather than
|
||
# a tool call.
|
||
- task_id: action_examples-messaging
|
||
n_examples: 3400
|
||
weight: 2.0
|
||
notes: |
|
||
100 records per messaging action (34 actions, 3,400 total). File:
|
||
data/synthesized/action_examples/messaging.jsonl. Multilingual
|
||
coverage zh/es/fr/ja/de/pt (5 per language per action); ~10%
|
||
subtle-null REPLY/IGNORE records. Slug matches the pack-time
|
||
`<parent_dir>-<stem>` derivation in scripts/pack_dataset.py so the
|
||
weight applies during sampling.
|
||
|
||
# Commerce / scheduling tool-call examples generated by
|
||
# scripts/synthesize_commerce_actions.py. Targets 18 actions across four
|
||
# plugins:
|
||
# plugin-shopify (5) MANAGE_SHOPIFY_CUSTOMERS, MANAGE_SHOPIFY_INVENTORY,
|
||
# MANAGE_SHOPIFY_ORDERS, MANAGE_SHOPIFY_PRODUCTS,
|
||
# SEARCH_SHOPIFY_STORE
|
||
# plugin-calendly (3) BOOK_CALENDLY_SLOT, CANCEL_CALENDLY_BOOKING,
|
||
# LIST_CALENDLY_EVENT_TYPES
|
||
# plugin-cron (5) CREATE_CRON, DELETE_CRON, LIST_CRONS, RUN_CRON,
|
||
# UPDATE_CRON
|
||
# plugin-commands (5) COMMANDS_LIST, HELP_COMMAND, MODELS_COMMAND,
|
||
# STATUS_COMMAND, STOP_COMMAND
|
||
# 100 records per action (1,800 total) with ~70% English / ~30%
|
||
# multilingual (zh/es/fr/ja/de/pt — 90 per language across actions, well
|
||
# above the >=3-per-language-per-action floor) and ~6% subtle-null REPLY
|
||
# records that emit a thought:/text: clarifying-question shape when the
|
||
# user message is too underspecified to resolve. Polymorphic
|
||
# MANAGE_SHOPIFY_* actions spread examples across documented sub-ops
|
||
# (list/search/get/create/update/delete/check/locations/adjust/status/
|
||
# fulfill/global). Argument schemas: most actions are parameter-free
|
||
# (arguments:{}); the four actions that expose `confirmed:boolean`
|
||
# (MANAGE_SHOPIFY_INVENTORY, MANAGE_SHOPIFY_ORDERS,
|
||
# MANAGE_SHOPIFY_PRODUCTS, CANCEL_CALENDLY_BOOKING) emit
|
||
# `confirmed: true` only when the user message clearly authorizes the
|
||
# write. 8 user phrasing styles (direct, formal, casual, expert-shorthand,
|
||
# voice-asr, distracted-rambling, broken-english, self-correcting) over
|
||
# 24 personas.
|
||
- task_id: action_examples-commerce
|
||
n_examples: 1800
|
||
weight: 2.0
|
||
notes: |
|
||
100 records per commerce / scheduling action (18 actions, 1,800
|
||
total). File: data/synthesized/action_examples/commerce.jsonl. Covers
|
||
plugin-shopify, plugin-calendly, plugin-cron, plugin-commands.
|
||
Multilingual coverage zh/es/fr/ja/de/pt (90 per language); ~6%
|
||
subtle-null REPLY records. Slug matches the pack-time
|
||
`<parent_dir>-<stem>` derivation in scripts/pack_dataset.py so the
|
||
weight applies during sampling.
|
||
|
||
# Translated stratified sample of train.jsonl across 8 languages, generated
|
||
# by scripts/translate_corpus.py using argos-translate (offline NMT).
|
||
# Action names, JSON keys, paths, URLs, and ALL_CAPS identifiers
|
||
# are masked before translation and restored after, so the structural
|
||
# contract of expectedResponse stays intact. Files live under
|
||
# data/synthesized/translated/<lang>.jsonl.
|
||
# 17 of the 200 sampled records exceed an 8 KB content cap and are skipped
|
||
# for translation runtime, leaving 183 records per language. Russian was
|
||
# cut short at 99 because the en->ru argos pack is ~3x slower than the
|
||
# others; see TRANSLATION_REPORT.md for details.
|
||
- task_id: translated-es
|
||
n_examples: 183
|
||
weight: 0.5
|
||
notes: Spanish translations of stratified train sample (183 records).
|
||
- task_id: translated-fr
|
||
n_examples: 183
|
||
weight: 0.5
|
||
notes: French translations of stratified train sample (183 records).
|
||
- task_id: translated-de
|
||
n_examples: 183
|
||
weight: 0.5
|
||
notes: German translations of stratified train sample (183 records).
|
||
- task_id: translated-pt
|
||
n_examples: 183
|
||
weight: 0.5
|
||
notes: Portuguese translations of stratified train sample (183 records).
|
||
- task_id: translated-zh
|
||
n_examples: 183
|
||
weight: 0.5
|
||
notes: Mandarin (Simplified) translations of stratified train sample (183 records).
|
||
- task_id: translated-ja
|
||
n_examples: 183
|
||
weight: 0.5
|
||
notes: Japanese translations of stratified train sample (183 records).
|
||
- task_id: translated-ko
|
||
n_examples: 183
|
||
weight: 0.5
|
||
notes: Korean translations of stratified train sample (183 records).
|
||
- task_id: translated-ru
|
||
n_examples: 99
|
||
weight: 0.5
|
||
notes: |
|
||
Russian translations of stratified train sample (99 records).
|
||
Capped early; en->ru argos pack runs at ~0.2 rec/s vs ~0.5 for others.
|
||
|
||
# Web3 plugin action examples — canonical tool_call records targeting
|
||
# the 9 web3 actions across plugin-evm and plugin-solana. Generated by
|
||
# scripts/synthesize_web3_actions.py using realistic crypto contexts
|
||
# (EVM addresses, Solana base58 keys, governance proposal ids, mainstream
|
||
# token symbols) across 22 personas, 22 styles, and 7 languages
|
||
# (en/zh/ja/es/fr/de/pt). 5-10% of records per action are subtle-null
|
||
# cases that emit a thought:/text: REPLY shape when a required field is
|
||
# missing from the user message. plugin-evm.TRANSFER and
|
||
# plugin-solana.TRANSFER are disambiguated via metadata.synth_plugin.
|
||
- task_id: action_examples-web3
|
||
n_examples: 900
|
||
weight: 2.0
|
||
notes: |
|
||
100 examples per action × 9 actions:
|
||
plugin-evm: CROSS_CHAIN_TRANSFER, GOV_EXECUTE, GOV_PROPOSE,
|
||
GOV_QUEUE, GOV_VOTE, SWAP, TRANSFER
|
||
plugin-solana: SWAP_SOLANA, TRANSFER
|
||
File: data/synthesized/action_examples/web3.jsonl.
|
||
|
||
# Under-represented core-prompt coverage generated by
|
||
# scripts/synthesize_core_prompts.py. 100 records per task across 8
|
||
# critically thin core prompts. ~70% English / ~30% multilingual
|
||
# (zh/es/fr/ja/de/pt, >=3 each), 8 styles, 30 personas. Memory-depth
|
||
# mix: 30% empty / 50% 1-2 turns / 20% 3 turns. The
|
||
# expectedResponse for each task matches its registry-v2 schema and
|
||
# is valid JSON. Files:
|
||
# data/synthesized/core_prompts/<task_id>.jsonl. Slugs match the
|
||
# pack-time `<parent_dir>-<stem>` derivation in scripts/pack_dataset.py
|
||
# so the per-entry weight applies during sampling.
|
||
- task_id: core_prompts-reflection
|
||
n_examples: 100
|
||
weight: 2.0
|
||
- task_id: core_prompts-reflection_evaluator
|
||
n_examples: 100
|
||
weight: 2.0
|
||
- task_id: core_prompts-extract_secrets
|
||
n_examples: 100
|
||
weight: 2.0
|
||
- task_id: core_prompts-choose_option
|
||
n_examples: 100
|
||
weight: 2.0
|
||
- task_id: core_prompts-message_classifier
|
||
n_examples: 100
|
||
weight: 2.0
|
||
- task_id: core_prompts-multi_step_decision
|
||
n_examples: 100
|
||
weight: 2.0
|
||
- task_id: core_prompts-add_contact
|
||
n_examples: 100
|
||
weight: 2.0
|
||
- task_id: core_prompts-should_follow_room
|
||
n_examples: 100
|
||
weight: 2.0
|