Reproduce, inspect, and extend
The site is a static project report with downloadable numerical evidence. Its primary pages work without access to the source repository. Source documents are rendered in the reference library; original Chinese engineering documents are labeled as such rather than silently machine-translated.
Rebuild the frozen analysis
From the repository root, with Python 3.10 or later:
python -m pip install -e ".[analysis,site]"
python scripts/analyze_training.py
python scripts/analyze_behavior.py
python scripts/analyze_wandb.py
python scripts/plot_training.py
python scripts/plot_behavior.py
python scripts/plot_wandb.py
python scripts/build_site.py
python scripts/check_publication.py
python -m http.server 8000 --bind 127.0.0.1 --directory _site
Open http://127.0.0.1:8000. Default aggregation reads committed numerical extracts and frozen gate artifacts; it does not query models or W&B. Behavior bootstrap uses 5,000 task resamples, while primary test intervals use the existing comparison implementation with 10,000. Each records its random-number convention.
Refresh source measurements
analyze_training.py --refresh-local reads local training logs, triage files, and evaluation traces. analyze_behavior.py --refresh-local --tokenizer /path/to/tokenizer.json additionally parses all saved turns and re-encodes training outputs. This requires the ignored runs/ directory and the appropriate tokenizer; it is separate from ordinary reproduction.
To refresh the four authorized W&B histories, install the optional wandb extra and run python scripts/fetch_wandb.py. It reads WANDB_API_KEY from the environment or project .env only in memory. The script exports numeric history and minimal source metadata; use scan_history() to avoid the default sampling of history().
Run the project
The repository contains the shopping environment, policy runtime, evaluation CLI, training adapters, Trace2Skill pipeline, configs, and frozen artifacts. Data corpora, model checkpoints, large raw traces, local service credentials, and machine caches are not included in the site bundle.
Start with the environment guide, then runtime and evaluation. The training guide covers data preparation, frozen curriculum checks, GPU topology, slime launching, failure analysis, bare validation, and online gates. Use a new output directory for new experiments so historical artifacts remain traceable.
Check the analysis contracts
The extraction tests distinguish a product visit from an option refresh, count protocol repair as a model turn but not an action, reject partial token denominators, keep technical failures separate, and verify task-level aggregation before pairing. Publication checks validate bilingual pages, local links and anchors, source hashes, figure XML, four-round W&B coverage, and agreement of reported test totals.
Data package
| Evidence | Download |
|---|---|
| Primary analysis and skill bank texts | analysis.json |
| Generated-group summaries | training_groups.csv |
| Primary rollout metrics | training_steps.csv |
| Per-task four-condition outcomes | test_task_outcomes.json |
| Gate version coefficients | chunk_contributions.csv |
| Behavioral episode / turn extracts | behavior report |
| Optimizer and systems history | W&B diagnostics |
| Input hashes | provenance.json |
Publication and reuse
GitHub Pages builds _site/ using an explicit asset allowlist; credentials, raw run directories, and corpus files are outside the package. The build contains a checksum manifest. Deployment is configured through the repository's Pages workflow and requires the repository/account to support the chosen visibility.
ShopSimulator retains upstream provenance and notices. The project does not invent a new license for upstream materials or imply that local experimental use grants unrestricted redistribution rights. See the provenance record and existing repository notices.