Skip to content

Repository files navigation

CommitBench

PPT -> https://docs.google.com/presentation/d/105Rw6dBRRDFlABEg7VL8iJBnOge10DoGh79Z1J3YwSE/edit?slide=id.g3ee20b0a688_2_2#slide=id.g3ee20b0a688_2_2

Local handoff/setup repo for the EffectBench-Ω minimal-plus experiment suite.

Current active plan: local/open-weight only. Bedrock/API runs are archived and should not be used for headline claims.

Historical Bedrock smoke notes exist in older runbooks, but the active PLAN.md path does not require Bedrock access.

Current Status

  • Main Bedrock experiments: not run.
  • Full local Qwen experiment grid: not run by request.
  • Local-open 21,504-row manifest: built at effectbench_omega/manifests/tasks_local_open.csv.
  • Selected-four local-open dry smoke: passed on 60 traces.
  • Historical live local model-call smoke: passed with qwen3_14b_awq_local on 1 trajectory, 0 failures.
  • Selected-four live model-call smokes: passed for Mistral-Small-3.2-24B, Qwen3.6-35B-A3B, Gemma-3-27B-it, and Llama-3.3-70B-AWQ.
  • Open model cache: Mistral-Small-3.2-24B, Qwen3.6-35B-A3B, Gemma-3-27B-it, and Llama-3.3-70B-AWQ are ready in the shared HF cache. Qwen3-14B-AWQ and Qwen3-30B-A3B are also cached but no longer in the locked headline grid.
  • Local vLLM endpoint: stopped after completed Stage 2b queue; no experiment server is currently required.
  • Upstream benchmark repos: cloned and pinned.
  • Headline local-only manifest: generated, 21,504 rows.
  • End-to-end workflow smoke: passed on local dry trajectories plus one live local Qwen14 call.
  • Final local verification pass: passed.
  • Native-source audit: passed; four benchmark families are source-backed, delegated-docs is explicitly controlled synthetic.
  • Stage 1 live per-model smoke: passed for all four locked local models, 12 traces/model, 0 failures/model, no-oracle 100%, local cost $0.
  • Stage 2 ordered local queue: complete. All four local slices finished, 21,504 traces total, 0 final failures, no-oracle 100%, local cost $0.
  • Model-controlled action-policy smoke: passed with hardened defaults. local_open_smoke_mc_default_20260622T204613Z ran 21 traces/model across all regimes and systems, 84 total traces, 0 failures, no-oracle 100%, local cost $0, JSON proposals 21/21 for every model, empty repair logs 21/21 for every model, proposal diversity gate passed.
  • Stage 2b model-controlled queue: complete as local_open_main_mc_postfix_20260622T214941Z; all four slices finished, 21,504 traces, 0 failures, no-oracle 100%, local cost $0.
  • Stage 3 merge: complete at effectbench_omega/outputs/main_mc_postfix_all_local/; 21,504 merged traces, 21,504 API log rows, 0 failures.
  • Stage 3 hardened offline suite: complete as stage3_hardened_main_mc_postfix_all_local_20260623T220316Z; 21,504 certificates, 5,149 strict-excess certificates, 0 unresolved abstraction warnings, no-oracle aggregate pass, 160 replay bundles checked with 0 failures.
  • Hardened metrics: complete. Projection baselines are data-derived, bootstrap uses paired/task-cluster/hierarchical resampling with 2,000 draws, CEGAR uses reduced abstract-state collision checks, and guard_tie explains PROJ_GUARD vs EFFECTGUARD.
  • Stage 3 figures and lite claim registry: complete. Aggregate writes PNG/PDF figures under effectbench_omega/figures/main_mc_postfix_all_local/ and a 50-row computed registry at effectbench_omega/metrics/claim_registry_main_mc_postfix_all_local.csv.
  • Queue defaults: Step 2b hardened path. run_local_open_queue.sh now defaults to main_mc_postfix, MODEL_CONTROLS_POLICY=1, MODEL_PROPOSAL_MODE=actions, TP=4 on GPUs 0,1,2,3.
  • Prompt fairness: locked. All four local models receive the same system instruction, task context, user-turn rendering, action enum, terminal_action requirement, temperature, and max token budget; only the structured-output transport differs by serving support.

Current Caveats

  • Qwen full slice has 42/5,376 audited repaired-fallback proposals. The primary run remains frozen; any same-prompt rerun of those rows should be reported as a sensitivity/ablation, not silently substituted.
  • Stage 3 projection/bootstrap/CEGAR are now replayable data-derived local audits. They are still simulator-local, not external human audits. CEGAR no-collision fields are a conservative-audit strength: the checker does not reject omissions without observed label-changing collisions; that means "not exercised by this scaffold," not "never future-relevant."
  • PROJ_GUARD and EFFECTGUARD are effectively tied in this local implementation: only 3/7,168 paired units differ in verifier verdict, all Mistral telecom confirm-target cases.
  • Active claims are local/open-weight only. Do not claim Bedrock/frontier leaderboard coverage from this run.

Implemented Features

Area Status Notes
Task families Working local adapters Retail, airline, telecom, delegated docs, and ToolSandbox/contract-style tasks.
Regimes Working FULL, CONCAT, SHARDED, SNOWBALL, REVISE, MEMORY_REVISE, ADV_EFFECT.
Manifest builder Working effectbench_omega/manifests/tasks_local_open.csv has the locked 21,504-row denominator.
Online systems Working local policies BASE, PROJ_GUARD, EFFECTGUARD.
Model proposal parser Working JSON, text-scan, LOW/HIGH legacy, explicit repaired fallback action proposals, required terminal_action, Mistral tool calls, and JSON-schema constrained proposals for Qwen/Llama/Gemma.
No-oracle guard Working Runtime forbidden-field checks plus audit reports.
Kernel verifier Working Emits certificates, frontiers, prune logs, rejected-abstraction table, and summary JSON.
Offline metrics/audits Working Projection loss, CEGAR, lattice sensitivity, bootstrap, cost audit, aggregate tables, figures, claim registry, replay bundles.
Local vLLM models Working Generic launcher supports the active local model keys; Mistral uses official Mistral tokenizer/config/tool parser.
Stage 1 live smokes Passed effectbench_omega/scripts/run_selected4_live_smokes.sh; four locked models, 48 total live model-call traces, 0 failures.
Stage 2 queue Complete effectbench_omega/scripts/run_local_open_queue.sh; four local slices complete.
Stage 2b model-controlled queue Complete main_mc_postfix; four local model slices complete, 21,504 total traces, 0 failures.
Stage 3 merge helper Complete effectbench_omega/scripts/merge_local_open_slices.py; merged output is effectbench_omega/outputs/main_mc_postfix_all_local/.
Stage 3 hardened offline suite Complete effectbench_omega/scripts/run_stage3_offline.sh; hardened projection/bootstrap/CEGAR/guard-tie outputs generated for main_mc_postfix_all_local.
Bedrock Archived Not part of the active local-only plan.

Key Outputs

  • Active runbook/checkpoint tracker: effectbench_omega/RUNBOOK_LOCAL_ONLY.md
  • Archived Bedrock-era runbook: effectbench_omega/RUNBOOK.md
  • Local-open model cache report: effectbench_omega/artifacts/local_open_model_cache.json
  • Local Qwen cache report: effectbench_omega/artifacts/local_qwen_cache.json
  • Repo/model pins: effectbench_omega/artifacts/repo_versions.json
  • Workflow smoke dry outputs: effectbench_omega/outputs/workflow_smoke_dry/
  • Workflow smoke local-Qwen outputs: effectbench_omega/outputs/workflow_smoke_local_qwen/
  • Selected-four dry outputs: effectbench_omega/outputs/selected4_dry/
  • Selected-four live smoke automation: effectbench_omega/scripts/run_selected4_live_smokes.sh
  • Selected-four live endpoint smoke reports: effectbench_omega/reports/live_smokes/
  • Selected-four Stage 1 live outputs: effectbench_omega/outputs/smoke_<model>/
  • Ordered Stage 2 queue runner: effectbench_omega/scripts/run_local_open_queue.sh
  • Ordered Stage 2 queue monitor: effectbench_omega/scripts/show_local_open_queue_status.sh
  • Pre-fix model-controlled balanced smoke outputs: effectbench_omega/outputs/smoke_mc_balanced_<model>/
  • Accepted default model-controlled smoke outputs: effectbench_omega/outputs/smoke_mc_default_<model>/
  • Recommended next full model-controlled outputs: effectbench_omega/outputs/main_mc_postfix_<model>/
  • Merged Stage 2b output: effectbench_omega/outputs/main_mc_postfix_all_local/
  • Stage 3 hardened job log: effectbench_omega/jobs/stage3_hardened_main_mc_postfix_all_local_20260623T220316Z/
  • Stage 3 main tables: effectbench_omega/tables/main_mc_postfix_all_local/
  • Stage 3 hardened projection table: effectbench_omega/tables/projection_loss_main_mc_postfix_all_local.csv
  • Stage 3 hardened uncertainty table: effectbench_omega/tables/uncertainty_main_mc_postfix_all_local.csv
  • Stage 3 hardened CEGAR table: effectbench_omega/tables/cegar_rejections_main_mc_postfix_all_local.csv
  • Stage 3 guard-tie report: effectbench_omega/tables/guard_tie_main_mc_postfix_all_local.md
  • Stage 3 figures: effectbench_omega/figures/main_mc_postfix_all_local/
  • Stage 3 witness bundles: effectbench_omega/witness_bundles/main_mc_postfix_all_local/
  • Stage 3 lite claim registry: effectbench_omega/metrics/claim_registry_main_mc_postfix_all_local.csv
  • Live Qwen14 smoke outputs: effectbench_omega/outputs/local_runner_qwen14_smoke/
  • Selected-four smoke certificates: effectbench_omega/outputs/smoke_<model>/kernel/certificates.parquet
  • Native source audit: effectbench_omega/reports/native_source_audit.md

See effectbench_omega/RUNBOOK_LOCAL_ONLY.md for staged commands, status, and the next experiment steps.

About

No description, website, or topics provided.

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages