chopratejas--headroom/headroom/evals

Headroom Evaluation Framework

Prove that compression preserves LLM accuracy through rigorous OSS benchmarks.

Results

Standard Benchmarks — "No Accuracy Loss"

Benchmark	Category	N	Baseline	Headroom	Delta
GSM8K	Math	100	0.870	0.870	0.000
TruthfulQA	Factual	100	0.530	0.560	+0.030

Compression Benchmarks — "Big Savings, Accuracy Preserved"

Benchmark	Category	N	Accuracy	Compression	Method
SQuAD v2	QA	100	97%	19%	Before/After
BFCL	Tool/Function	100	97%	32%	LLM-as-Judge
Tool Outputs (built-in)	Agent	8	100%	20%	Before/After + Proxy
CCR Needle Retention	Lossless	50	100%	77%	Exact Match

Model: gpt-4o-mini | Suite cost: ~$3 | Duration: ~15 min

Installation

pip install "headroom-ai[all]"    # Everything including evals (recommended)
pip install "headroom-ai[evals]"  # Evaluation framework only

Quick Start

Run the Evaluation Suite

# Quick smoke test (8 cases, ~10s)
python -m headroom.evals quick -n 8 --provider openai --model gpt-4o-mini

# Full Tier 1 suite (~$3, ~15 min) — requires proxy running
python -m headroom.evals suite --tier 1 -o eval_results/

# Extended suite (Tiers 1+2, ~$8, ~1 hr)
python -m headroom.evals suite --tier 2 -o eval_results/

# CI mode — exit 1 on any regression
python -m headroom.evals suite --tier 1 --ci

# List all available datasets
python -m headroom.evals list

Running with the Proxy (Recommended)

For the most accurate evaluation, run through the Headroom proxy which provides the full stack: compression + CCR retrieval + cache alignment.

# Terminal 1: Start the proxy
headroom proxy --port 8787

# Terminal 2: Run evals (auto-detects proxy)
python -m headroom.evals suite --tier 1 -o eval_results/

Without the proxy, the eval runner falls back to local compression only (no CCR).

Python API

from headroom.evals.suite_runner import SuiteRunner
from headroom.evals.reports.report_card import save_reports

# Run Tier 1 suite
runner = SuiteRunner(model="gpt-4o-mini", tiers=[1])
result = runner.run()

# Save Markdown, JSON, and HTML reports
save_reports(result, "eval_results/")

Session Probes (real recorded sessions)

The benchmark suites above prove the compressors in vitro. Session probes measure what compression removed from YOUR real sessions, offline and with no LLM or API key:

# 1. Record: run the proxy with recording enabled (opt-in; recordings contain
#    full conversation content in plaintext and stay on this machine)
HEADROOM_PROBE_RECORD_DIR=~/.headroom/probe-recordings headroom proxy start

# 2. Use your agent through the proxy as normal, then score retention:
headroom evals probes --recordings ~/.headroom/probe-recordings

Probe targets are extracted from original tool results across three dimensions — exact numerics, artifact trail (paths/URLs/hashes), error evidence — and classified as retained (verbatim or surviving a format conversion), recoverable (behind a CCR retrieval marker), or lost. The report buckets retention by compression ratio and groups it per transform. Use --json-output report.json for machine-readable results.

Evaluation Tiers

Tier 1: Core Report Card (~$3, ~15 min)

Benchmark	Runner	What It Tests
GSM8K	lm-eval harness	Math reasoning accuracy
TruthfulQA	lm-eval harness	Factual accuracy
MMLU	lm-eval harness	57-subject knowledge
ARC-Challenge	lm-eval harness	Science reasoning
HumanEval	lm-eval harness	Code generation
SQuAD v2	Before/After	Reading comprehension with compression
BFCL	LLM-as-Judge	Function calling with compressed schemas
Tool Outputs	Before/After + Proxy	Agent tool output compression
CCR Needle Retention	Compression-only	Lossless anomaly preservation

Tier 2: Extended (~$5 more, ~30 min)

Benchmark	Runner	What It Tests
HotpotQA	Before/After	Multi-hop QA with compressed passages
MS MARCO	Before/After	RAG with compressed search results
CodeSearchNet	Before/After	Code understanding after compression
Info Retention	Compression-only	Probe fact survival in compressed output

Tier 3: Deep Dive (~$9 more, ~45 min)

Benchmark	Runner	What It Tests
HellaSwag	lm-eval harness	Commonsense reasoning
NarrativeQA	Before/After	Long narrative comprehension
TriviaQA	Before/After	Factoid QA at scale

Evaluation Methods

Before/After (Default)

Compares LLM responses on original vs. compressed context:

Original Context ──► LLM ──► Response A
                                         ├─► Compare (F1, semantic sim, GT match)
Compressed Context ──► LLM ──► Response B

When the Headroom proxy is running, the "compressed" path goes through the full stack (compression + CCR tool injection + cache alignment), which is the real production experience.

LLM-as-Judge (for BFCL, tool use)

Uses an LLM judge to compare the compressed response against ground truth semantically. This handles cases where the same correct answer can be expressed in different formats (function call JSON vs natural language computation).

Compressed Context ──► LLM ──► Response ──► LLM Judge ──► Score 1-5
                                                  ▲
                                          Ground Truth

Score >= 3 ("partially correct or better") = PASS. This means the compressed context preserved enough information for the LLM to reach the right answer.

Compression-Only (Zero Cost)

Tests compression quality without any LLM API calls:

• CCR Needle Retention: Compress JSON arrays, verify errors/anomalies survive
• Information Retention: Compress and check if probe facts are preserved

Available Datasets

RAG / Retrieval

Dataset	Description	Default N
hotpotqa	Multi-hop QA over multiple Wikipedia passages	100
natural_questions	Real Google search questions with Wikipedia answers	100
triviaqa	Large-scale trivia QA with evidence documents	100
msmarco	Real Bing search queries with relevant passages	100
squad	SQuAD v2 reading comprehension	100

Tool Use

Dataset	Description	Default N
bfcl	Berkeley Function Calling Leaderboard	100
toolbench	Real-world API tool usage scenarios	100
tool_outputs	Built-in realistic tool outputs (JSON, logs, etc.)	8

Long Context

Dataset	Description	Default N
longbench	Long context understanding (4K-128K tokens)	50
narrativeqa	Story comprehension	100

Code

Dataset	Description	Default N
codesearchnet	Code snippets with descriptions	100
humaneval	Programming problems (OpenAI)	164

Metrics

Metric	Description	Pass Threshold
F1 Score	Token overlap between responses	> 0.7
Semantic Similarity	Embedding cosine similarity	> 0.85
Ground Truth Match	Answer present in response	True
LLM Judge Score	1-5 semantic correctness scale	>= 3
Accuracy Preserved	Any of the above passes	True

CI/CD Integration

# .github/workflows/eval.yml
name: Evaluation Suite

on:
  pull_request:
    paths: ['headroom/transforms/**', 'headroom/evals/**']
  schedule:
    - cron: '0 6 * * 1'  # Weekly

jobs:
  smoke-test:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: actions/setup-python@v5
        with: { python-version: "3.11" }
      - run: pip install -e ".[all]"
      - name: CCR Round-trip (zero cost)
        run: |
          python -c "
          from headroom.evals.runners.compression_only import CompressionOnlyRunner
          r = CompressionOnlyRunner()
          result = r.evaluate_ccr_lossless(r.generate_ccr_test_cases(50))
          assert result.passed, f'CCR failures: {result.errors}'
          print(f'CCR: {result.passed_cases}/{result.total_cases} PASS')
          "
      - name: Quick eval
        run: python -m headroom.evals quick -n 8 --provider openai --model gpt-4o-mini
        env:
          OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}

Environment Variables

Set in .env at project root (auto-loaded by the suite runner):

OPENAI_API_KEY=sk-...      # Required for OpenAI models
ANTHROPIC_API_KEY=sk-ant-... # Required for Anthropic models

Architecture

headroom/evals/
├── __init__.py              # Public API
├── __main__.py              # CLI (quick, list, benchmark, suite, report)
├── core.py                  # EvalCase, EvalResult, EvalSuite
├── datasets.py              # 12 dataset loaders (HuggingFace + built-in)
├── metrics.py               # F1, semantic similarity, ROUGE-L, BLEU
├── cost_tracker.py          # API spend tracking + budget enforcement
├── suite_runner.py          # Tiered suite orchestrator (16 benchmarks)
├── comprehensive_benchmark.py  # EleutherAI lm-eval harness wrapper
├── runners/
│   ├── before_after.py      # Before/After + LLM-as-Judge + proxy support
│   └── compression_only.py  # Zero-cost CCR + info retention evals
├── reports/
│   └── report_card.py       # Markdown, JSON, HTML report generation
└── memory/
    ├── judge.py             # LLM-as-judge (OpenAI, Anthropic, LiteLLM)
    └── runner*.py           # Memory-specific evaluation