AI/Tasks/CurrentTask.txt
|
Resume the recommendation micro-pass dataset workflow and produce verifiable build/test artifacts.
Repository root: C:\repos\TechToolbox Primary scope: C:\repos\TechToolbox\ModelTraining Objective: - Build dataset artifacts for the recommendation micro pass. - Run a bounded training pass with repo-root virtual environment resolution. - Run evaluation and report score deltas from generated artifacts only. Critical safety constraints: - Never write to, replace, or mutate any file under C:\repos\TechToolbox\.venv\. - Never use WRITE-FILE or REPLACE-IN-FILE on executable files. - Do not claim training or evaluation success without a completed output artifact. - If any command fails, capture exact command, exact stderr/exception, and stop after one focused retry. Execution plan (in order): 1) Environment precheck: - Verify interpreter path and version: - .\\.venv\\Scripts\\python.exe --version - Verify required files exist: - ModelTraining\Build-AgentControlRecommendationMicroPass2Dataset.ps1 - ModelTraining\Train-QwenLoRA.ps1 - ModelTraining\Evaluate-DecisionPack.py - ModelTraining\decision_eval_pack_extended.json 2) Rebuild deterministic pass-2 dataset: - Run: - pwsh -NoProfile -File .\ModelTraining\Build-AgentControlRecommendationMicroPass2Dataset.ps1 - Capture sample count and output path. 3) Rebuild train/validation jsonl for pass-2: - Confirm these files exist and are non-empty: - ModelTraining\train-recommendation-micro-pass2.jsonl - ModelTraining\validation-recommendation-micro-pass2.jsonl - If missing, identify exact missing generation step and run only that step. 4) Bounded training run (no blind long run): - Run Train-QwenLoRA with explicit pass-2 files and a dedicated output dir: - .\ModelTraining\Train-QwenLoRA.ps1 -TrainJsonl .\ModelTraining\train-recommendation-micro-pass2.jsonl -ValidationJsonl .\ModelTraining\validation-recommendation-micro-pass2.jsonl -OutputDir .\ModelTraining\lora-recommendation-micro-pass3\final - Report exit code and final adapter output path. 5) Evaluation run: - Run evaluator with: - base model: Qwen/Qwen2.5-7B-Instruct - adapter dir: .\ModelTraining\lora-recommendation-micro-pass3\final - prompt pack: .\ModelTraining\decision_eval_pack_extended.json - output report: .\ModelTraining\decision_eval_report_extended_recommendation_micro_pass3.json - Use repo-root venv python directly. 6) Compare to current anchor: - Compare new report overall pass metrics to: - ModelTraining\decision_eval_report_extended.json - ModelTraining\decision_eval_report_recommendation_micro.json - State clearly: improved, tied, or regressed. Output format: 1) Commands executed. 2) Artifacts generated (paths + existence + size). 3) Training result (exit code, adapter output path). 4) Evaluation result (overall pass counts and delta). 5) Final verdict: improved, tied, regressed, or blocked. 6) If blocked, provide exactly one next command to isolate the blocker. Quality bar: - Use only observed command output and generated artifacts as evidence. - No speculation words when evidence exists. - Keep retries minimal and targeted. |