You need to agree to share your contact information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
Research experiment artifacts. Access is reviewed manually.
Log in or Sign Up to review the conditions and access this dataset content.
Berkeley WebAgent Experiment Artifacts
Native GEPA, CLUE and ACE experiment logs and available actor trajectory evidence. Files require manual access approval. Request access with your Hugging Face account.
Results and complete evidence snapshot — September 23, 2026
- Combined experiment summary: GEPA, CLUE and ACE; method × site for WebArena.
- Non-WebArena repeat scores and variance, including completed ACE ≤50k-token evaluations.
- WebArena / GoBrowse progress and audited final repeats.
- Systematic raw evidence snapshot: 590,861 file paths, source/export checksums, original and invalidated attempts, checkpoints and restoration script.
This is a timestamped snapshot. WebArena training continues; incomplete groups are explicitly marked. Original ACE context-limit failures and separate reevaluations remain distinct.
Results — September 21, 2026
- Non-WebArena results: training, validation and five-repeat held-out tables/plots.
- Every iteration: training and validation by iteration; separate ACE training windows.
- WebArena speed: component benchmarks, direct commit comparison and measurement limits.
- CSV/JSON, PNG/SVG and provenance.
Previous trace snapshot — September 21
September 21 trace snapshot: stored native trajectories for nine completed non-WebArena runs and four partial WebArena site runs, plus available full ACE event journals. Coverage, failures, source/export hashes and archive links are in the manifest.
Layout
logs/<snapshot>/: requested launch/optimization/LM-call logs.traces/<run-id>/<snapshot>/: compressed episode shards with export manifests and checksums.index/<snapshot>/manifest.json: run coverage, artifact hashes, sizes and missing evidence.
GEPA/CLUE launch.log aliases the native launcher.log. ACE AppWorld retains native per-task and optimizer-phase lm_calls.jsonl paths. ACE tau-bench has separate reflector/curator logs. Historical ACE ALFWorld optimizer originals were not saved.
Only stored evidence is exported. Missing past traces are not reconstructed. Runs may still be in progress. Trace shards contain completed recordings at the snapshot cutoff; a completed recording does not imply task success or a finished run. Original files remain unchanged. Credentials are redacted from export copies; source and exported hashes are distinguished. Returned reasoning is included; provider-internal reasoning unavailable through the API cannot be recorded.
These are research artifacts, not a benchmark split or a complete matched-model comparison. Consult per-run settings, attempts and coverage before interpreting results.
- Downloads last month
- 77