New DAG (`ml/pipelines/bench_dag.py`) with three linked tasks:
1. collect.py — generates candidates, logs to MLflow
2. export_for_judge — exports pending runs for Claude Code scoring
3. compare — generates leaderboard by (model, prompt) cell
Config via dag_run.conf supports all collect.py options (models, prompts,
n_tips, n_scenarios, temperature, experiment name, max_model_b).
New admin API endpoints (`services/api/src/routes/bench.ts`):
- GET /api/bench/experiments — list tip-bench-* experiments
- POST /api/bench/run — trigger DAG with custom config
- GET /api/bench/runs/:experiment — list runs in experiment
- GET /api/bench/leaderboard/:experiment — leaderboard by (model, prompt)
All endpoints require admin auth. Human judge (Claude Code) scores are
applied manually post-export; future enhancement: add webhook to DAG.
Admin UI can now trigger and monitor benchmarks from a dashboard panel.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Combines model evaluation (#93) and prompt A/B testing (#95) into one
experiment. Evaluates all (model × prompt × scenario) cells on the same
fixed contexts so quality differences are attributable.
Architecture:
- Phase A (collect.py): generates candidates per cell, logs to MLflow
with judge_pending=true. Rejects models >4B, uses keep_alive=0 for
RAM safety (no concurrent model weights in VRAM).
- Phase B (judge_cli.py): exports pending runs as JSON for Claude Code
to score per the rubric, then applies scores back to MLflow.
- Phase C (compare.py): leaderboard by (model, prompt) cell.
Rubric (tip-v1) defines 1–5 scales for relevance, actionability, tone,
plus format_ok and overlong flags. Composite = rel + act + tone +
2×format_ok − overlong. Rubric is self-describing and persisted in every
run so judges use consistent criteria across sessions.
Artifacts (prompts, candidates, raw responses) stored as MLflow tags
because the server uses a file:// backend not accessible via REST. Full
artifacts accessible in MLflow UI → run → Tags section.
Tested end-to-end on local machine:
- 4 models (qwen2.5:0.5b/1.5b, gemma3:1b, llama3.2:3b) ≤4B
- 3 prompts (v1, v2-mentor, v3-few-shot)
- 4 scenarios (4 personas × 2 time-slots)
- 48 cells total, all judged and ranked
Winner: qwen2.5:1.5b × v3-few-shot (composite=12.75).
Ready for integration into Airflow prompt_ab_eval DAG and admin UI.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Ship the scaffolding for #99 (phase B.3 of #81):
- ml/serving: add /score/egreedy/v2, /reward/egreedy/v2, /stats/egreedy/v2
endpoints (D=12). New feature dims: completion/dismiss rates, mean dwell
(clipped 10min), preferred-hour alignment (cosine, 1-dim), tip volume (log).
Separate state file per user (_egreedy_v2.json). /reset clears v2 state too.
- ADR-0012: documents D=7→12 dimension change, normalization choices, shadow
rollout protocol, and promotion gate (offline sim win per ADR-0002).
- recommender.ts: register egreedy-v2-shadow in shadow-policy map (disabled by
default). When enabled, calls /score/egreedy/v2 fire-and-forget and publishes
shadow:egreedy-v2-shadow serve signal. No reward to shadow — sim is the gate.
- sim runner/personas: personas carry synthetic profile_features per persona;
_call_score/_call_reward thread profile_features through (None-safe for v1/linucb).
- 18 new Python tests; all 56 Python + 170 TS tests pass.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
