- Restructure train_notebook.ipynb so the VM (sandbox) lane is the default
judges hit by running cells top-to-bottom; Mock-env path stays as commented
optional lines. Sandbox cluster prereq is moved into a markdown cell above
Step 4. Existing executed outputs (training run, sandbox compare, mentor
curves) are preserved untouched.
- Consolidate the three READMEs (root, sandbox/, demo/results/) into a single
root README.md that documents Theme #3.1 alignment, the 3 modes (Ticket /
Incident-Mock / Incident-VM), the train-then-evaluate pipeline, and the
artifacts/ vs results/ split.
- Add Blog.md as a team-voiced narrative covering theme alignment plus the
baseline (untrained) vs trained (RL-fine-tuned) comparison.
- Promote curated evaluation snapshots into results/{simple,sandbox,training}/
so the submission carries the actual reports + reward curves.
- Remove deprecated demo/ assets (video script, slides, mini blog, etc.) and
the standalone sandbox/ + demo/results/ R…
Add env.step safety net so checkpoint hallucinations never crash evaluation.
Wrap selector and env.step in run_episode with heuristic fallback and concise step-error logging. Out-of-data customer/service/runbook ids now retry safely instead of crashing the run.
Made-with: Cursor
Harden checkpoint sanitizer against all sandbox/sim edge cases.
Validate per-action literals, anchor service_name and runbook_id to env-provided values, default missing required strings, and wrap selector and sanitizer in try/except so evaluation never crashes on malformed checkpoint output.
Made-with: Cursor
Fix HF Space no-docker sandbox launcher shutdown compatibility.
Use uvicorn Server boolean shutdown flags instead of event-style setters so launch_no_docker works on current uvicorn versions in Spaces.
Made-with: Cursor
Finalize checkpoint-first evaluation and sandbox submission workflow.
Default compare runs now use trained checkpoints with explicit heuristic fallback, add no-Docker sandbox launch for HF Spaces, and align notebook/docs/runbook around mentor-ready stage curves and submission checks.
Made-with: Cursor
Improve submission docs and training run transparency.
Refresh README/demo messaging around policy provenance, align notebook and compare-eval outputs to artifacts/eval, and remove checkpoint eval timeout to avoid long-load truncation.
Made-with: Cursor
Improve checkpoint policy quality and environment robustness.
Run trained checkpoint evaluation in a clean subprocess to avoid Unsloth patch conflicts, and sanitize checkpoint actions so invalid outputs fall back to trained_heuristic only when required.
Made-with: Cursor
audit pass: hackathon-ready polish and truthful reporting
Surface policy_used and episodes_per_difficulty in EvaluationReport so artifacts unambiguously distinguish checkpoint vs heuristic-fallback runs. Replace hardcoded behavior_examples in train.py with real behavior_diffs derived from actual baseline/trained metrics. Rewrite README around the theme 3.1 judging criteria with a TL;DR, criterion mapping, and realistic time estimates. Split training notebook into quick (5a) and full (5b) modes, use unbuffered logs, and inspect policy_used in the results cell. Harden .gitignore against ephemeral venv dirs.
Made-with: Cursor
Penalize multi-JSON and trailing prose outputs, add near-cap completion penalties, and expose new diagnostics so training resists verbose reward-hacking and converges toward single compact JSON actions.
Made-with: Cursor
harden reward callback and add training diagnostics
Factor per-sample scoring into a helper with defensive try/except, coerce non-string inputs, and log length/unterminated stats plus a sample completion so stalls or reward-hacking patterns are visible during GRPO training.
Made-with: Cursor
Reduce GRPO learning rate and strengthen penalties for unterminated/overlong outputs so training is less prone to clipping saturation and verbose reward hacking.
Made-with: Cursor
Add honest results/reproduce sections and judge quick-start in README, ship checkpoint-based evaluation policy and compare flag in evaluate.py, wire trained adapter eval into train.py with graceful fallback, refresh blog/video/pitch narratives, and update notebook with clean install plus checkpoint compare and two-seed reproducibility cells.
Made-with: Cursor
Harden training signal and clean generated artifacts.
Improve GRPO robustness with tolerant JSON parsing, action-keyed reward shaping, parse-rate diagnostics, and shorter completion defaults; add raw reward reporting/plotting in evaluation for transparent baseline behavior; and remove redundant generated result snapshots.
Made-with: Cursor
Handle TRL/Unsloth trainer argument-name drift by mapping config-vs-args and related reward/dataset keyword variants so Colab runs do not crash on wrapped trainer signatures.
Made-with: Cursor
Fix TRL GRPO compatibility and Unsloth import ordering.
Handle GRPOConfig and GRPOTrainer signature drift across TRL versions so Colab runs do not fail on max_new_tokens/tokenizer argument changes, and import Unsloth before TRL stack to preserve optimization patches.
Made-with: Cursor
Harden hackathon readiness and surface reproducible evidence.
Align manifest and API docs with full environment lifecycle, improve training reproducibility and reward shaping, clarify heuristic-vs-trained evaluation semantics, and add tracked evaluation artifacts/checklist items for safer judging and reruns.
Made-with: Cursor
Add a dedicated notebook cell that securely captures the Hugging Face token without hardcoding it in the repo, and propagate HF_TOKEN/HUGGINGFACE_HUB_TOKEN to the training subprocess so model downloads avoid anonymous rate limiting and partial file fetch failures.
Made-with: Cursor
Fall back to transformers+peft when unsloth fails to import for any reason (including shared-GPU OOM during module initialization), and make the training notebook cell pick the GPU with the most free memory and enable the fallback flag so training proceeds.
Made-with: Cursor
Avoid importing unsloth directly in the notebook sanity cell to prevent unsloth_zoo CUDA OOM during startup, while still validating package versions, CUDA visibility, and dry-run training pipeline wiring.
Made-with: Cursor
Remove notebook installation of unavailable bitsandbytes builds, force Step 3 to bypass bitsandbytes integration, and update train.py to automatically fall back to non-4bit model loading when bitsandbytes cannot be imported.
Made-with: Cursor
upgrade bitsandbytes and wire torch CUDA runtime for DGX
Use a recent bitsandbytes release that leverages torch's bundled CUDA runtime and add torch lib directory to LD_LIBRARY_PATH in the sanity check cell so unsloth and bitsandbytes initialize without a system CUDA installation.
Made-with: Cursor
repair broken aiohttp/charset stack for unsloth imports
Prevent pyarrow and sentencepiece source builds by installing them with conda, use wheel-only pip installs for the core stack, and install unsloth/unsloth_zoo with no dependency re-resolution.
Made-with: Cursor
Update install cells to avoid pyarrow source builds and xformers compilation on DGX, remove mergekit-related click/pydantic conflicts, and force unsloth xformers fallback during sanity checks.
Made-with: Cursor
Use a DGX-safe install flow in the training notebook that installs pyarrow and datasets via conda, pins torch to 2.6, and installs xformers/unsloth in a way that avoids incompatible dependency resolution and source builds.
Made-with: Cursor
disable external report logging by default in GRPO training
Installs mergekit in Colab notebook and updates the training stack error hint so GRPO training no longer fails at import time with missing mergekit module.
Made-with: Cursor
fix UI button initialization regression on Spaces
Escapes newline regex in parseListInput so the served inline script is valid JavaScript; this restores Reset/Get state/Execute step button handlers in the web UI.
Made-with: Cursor
Replaces Windows-only absolute fixture paths with repository-relative paths so runbook and incident history tests pass consistently on Linux GitHub runners.
Made-with: Cursor
updated EICC v2 environment, APIs, and training pipeline
Adds the enterprise incident command center world model, expanded action/observation space, deterministic incident grading, production-hardening API updates, and complete training/evaluation artifacts for submission readiness.
Made-with: Cursor
