Mistral Model Ft Playbook

Use this skill to turn Mistral model experimentation into a repeatable production workflow.

Workflow

1. Choose model and endpoint strategy.

• Read references/mistral-official-models-and-customization.md.
• Prefer dated model IDs for stability.
• Choose API-hosted model vs open weights based on latency, cost, and deployment constraints.

2. Decide customization depth.

• Start with system prompt + eval.
• Escalate to fine-tuning only when prompting fails on format, style, or domain consistency.
• Use classifier/moderation layer as a separate safety control, not as a replacement for task tuning.

3. Normalize dataset format before training.

• Prefer {prompt, completion} when training strict output format tasks.
• Keep training prompt and inference prompt text aligned exactly.
• Keep evaluation split isolated from training split.

4. Run SFT with reproducible configuration.

• Log exact base model ID, dataset version, LoRA hyperparameters, and max sequence length.
• Treat adapter as default output artifact.
• Track job IDs and links for every run.

5. Validate generation quality with explicit parsers.

• Run inference checks on multiple fixed test cases.
• Enforce exact key set and value constraints.
• Fail fast when parse errors occur (no_json_start, unterminated_json, type/range mismatch).

6. Decide distribution format.

• Ship adapter when consumers can load base+adapter.
• Merge to full model only for easier deployment compatibility.
• Remember that merge changes packaging, not learned behavior quality.

Mistral-Specific Guardrails

• Match tokenizer/runtime artifacts to the serving path (tekken.json, params.json, chat_template.jinja, processor_config.json when applicable).
• For Ministral-3 style HF models, apply LoRA to the language model module expected by runtime.
• If PEFT integration fails at generation time, patch or ensure prepare_inputs_for_generation compatibility before wrapping.
• Preserve EOS handling explicitly in training and generation configuration for format-sensitive tasks.

This Repository Mapping

• Training: scripts/hf/train_sft_request_to_hidden_lm.py
• Dataset conversion: scripts/ft/convert_to_prompt_completion_dataset.py
• JSON debug inference: scripts/hf/debug_full_model_json_inference.py
• Adapter merge/upload: scripts/hf/merge_and_upload_full_model.py
• Job history: artifacts/hf_jobs/submissions.jsonl

References

• Official model/customization guidance: references/mistral-official-models-and-customization.md