elizaos--eliza/packages/benchmarks/abliteration-robustness/CLAUDE.md

Abliteration Robustness — Agent Guide

Over-refusal benchmark for abliterated model variants. Feeds the same harmless prompt set used by training/scripts/training/abliterate.py (default HuggingFace mlabonne/harmless_alpaca) and measures 1 - refusal_rate. Registered in the suite as abliteration-robustness.

Run

# Direct (vllm provider, adjust --base-url and --model as needed)
python -m benchmarks.abliteration-robustness.cli \
    --provider vllm \
    --model eliza-1-9b-abliterated \
    --base-url http://127.0.0.1:8001/v1 \
    --out /tmp/abliteration-robustness-out

# Through the suite orchestrator (resolves provider/model, stores results)
python -m benchmarks.orchestrator run \
    --benchmarks abliteration-robustness \
    --provider vllm \
    --model eliza-1-9b-abliterated

Override the dataset via --extra '{"dataset": "mlabonne/harmless_alpaca"}' or with a local JSONL via --extra '{"dataset_path": "/path/to/harmless.jsonl"}'.

Smoke test (no API keys)

python -m benchmarks.abliteration-robustness.cli \
    --provider mock \
    --model dummy \
    --out /tmp/abliteration-robustness-smoke

The mock provider uses three embedded harmless prompts and returns a canned helpful reply; no network or GPU required.

Test the harness

pytest abliteration-robustness/tests/ -v

Layout

Path	Role
cli.py	CLI entrypoint and all benchmark logic
__init__.py	Package marker
tests/test_abliteration_cli.py	pytest suite for argparser, harness dispatch, and refusal detector

Notes

• Results write to <out>/abliteration-robustness-results.json (path set by --out).
• Scored by _score_from_abliteration_robustness_json in registry/scores.py.
• Score = 1 - refusal_rate; higher is better (1.0 = no spurious refusals).
• Pair with an abliterated model variant; base models should also score well since the prompt set is intentionally benign.
• Full background: README.md.

⛔ NON-NEGOTIABLE — evidence, trajectories & real end-to-end tests

The binding, repo-wide standard is AGENTS.md. Read it. Nothing in this package is done until it is proven done — a reviewer must confirm it works without reading the code, from the artifacts you attach. This applies to every feature, fix, refactor, and chore here. "Tests pass" is not proof; "CI is green" is not proof.

• Record AND read model trajectories. Capture the actual inputs and outputs of the model from a live LLM — not the deterministic proxy, not a mock: the prompt, the providers/context, the raw model output, every tool/action call, and the result. Then open the trajectory and review it by hand. A captured-but-unread trajectory is not evidence (packages/scenario-runner/bin/eliza-scenarios run <scenario> --report <out>).
• Real, full-featured E2E — no larp. Every feature ships detailed end-to-end tests that drive the real path end to end. Not the happy "front door" only: cover error paths, edge/empty/invalid input, concurrency, roles/permissions, and adversarial input. A test that asserts against a mock/stub/fixture standing in for the thing under test does not count. If the real model/device/chain/connector/account is hard to reach, make it reachable — that is the work, not an excuse to mock. If the existing tests here are shallow or mocked, fixing them is part of your change.
• Screenshots + logs at every phase, plus a complete walkthrough video/run-through of the entire feature or view, start to finish (bun run test:e2e:record).
• Manually review every artifact the change touches — never just the green check: client logs (console + network), server logs ([ClassName] …), the model trajectories in and out, before/after full-page screenshots, and the domain artifacts listed below for this package.
• No residuals. No shortcuts. The goal is not "done" — it is everything done. Clear every blocker by the hard path: build the real architecture, stand up the real model/device/service, actually test it. Never leave a TODO, a stub, a stepping-stone, or a "follow-up." When unsure, research thoroughly, weigh the options, and ship the best, highest-effort, production-ready version. Keep going until every possibility is exhausted.

Artifacts → attached inline in the PR (MP4 video, JPG screenshots, logs in <details>); attach each evidence type or explicitly mark it N/A with a reason — never leave it blank. If develop moved and changed behavior, re-capture evidence; stale proof is worse than none.

Capture & manually review for this package — benchmark / eval suite:

• A real-model run (not the mock/smoke fixture) producing the score-report JSON, with the numbers inspected and the provider/model recorded.
• The per-item trajectories the harness captured, spot-reviewed for correctness — a green harness run over mock fixtures is not a result.
• The provider matrix actually exercised, and the scoring math validated against a known case.
• Failure / timeout / partial-output handling in the harness itself.