Mistral Hidden Params Ft

Use this skill to run the full loop from dataset preparation to deployed model validation.

Key Files

• scripts/ft/build_train_valid_test.mjs
• scripts/ft/convert_to_prompt_completion_dataset.py
• scripts/hf/train_sft_request_to_hidden_lm.py
• scripts/hf/debug_full_model_json_inference.py
• scripts/hf/merge_and_upload_full_model.py
• artifacts/hf_jobs/submissions.jsonl

Workflow

1. Rebuild split quality first.

• Use train-only data and regenerate train/validation/test before training when existing validation/test quality is low.

2. Normalize data to prompt-completion.

• Ensure each sample is {prompt, completion} with completion as strict JSON text.
• Keep the inference prompt text exactly aligned between conversion, training, and debug scripts.

3. Launch TRL SFT job on Hugging Face Jobs.

• Train with scripts/hf/train_sft_request_to_hidden_lm.py.
• Record job IDs and model IDs in artifacts/hf_jobs/submissions.jsonl.

4. Validate JSON output behavior with real generations.

• Run scripts/hf/debug_full_model_json_inference.py against either:
• Base + adapter mode (--base-model-id + --adapter-model-id)
• Full merged model mode (--model-id)
• Report valid_cases=X/Y and include representative malformed output when failures occur.

5. Merge adapter into full model only when distribution simplicity is needed.

• Use scripts/hf/merge_and_upload_full_model.py.
• Explain that merge does not inherently improve prediction quality; it simplifies inference deployment.

6. Report status with concrete links.

• Return Hugging Face job URL(s), output model repo URL(s), and whether each repo is adapter-only or full.

Guardrails

• Keep prompt wording identical across train/inference unless explicitly requested.
• Keep target keys exactly: energy, warmth, brightness, acousticness, complexity, nostalgia.
• Fail evaluation when output is not valid JSON object with exactly the required keys and integer range 0..10.
• Prefer evidence from actual job logs over assumptions.