Haruk1y
@Haruk1y
Public Skills
strict-json-output-hardening
by Haruk1y
Improve strict JSON generation reliability for fine-tuned language models using parser-based evaluation, prompt alignment, and targeted retraining loops. Use when outputs are malformed, have extra text, or violate schema/range constraints.
mistral-packaging-compat-check
by Haruk1y
Validate compatibility between Mistral model packaging format and inference path, including adapter versus merged full-model loading. Use when merged models fail generation or runtime/tokenizer artifacts are inconsistent.
wandb-weave-ft-retrospective
by Haruk1y
Analyze W&B, Weave, and local fine-tuning evaluation artifacts, then produce a concrete next-run improvement plan with data, prompt, and training actions. Use after each SFT or eval cycle.
hf-job-ops-playbook
by Haruk1y
Operate Hugging Face Jobs reliably for training and evaluation in this repository, with reproducible submission records, monitoring, and retry flow. Use when submitting, monitoring, or triaging HF Jobs.
mistral-model-ft-playbook
by Haruk1y
Build, fine-tune, evaluate, and ship Mistral-family models with Hugging Face and PEFT, with emphasis on strict JSON outputs and model-format compatibility. Use when choosing Mistral model variants, deciding prompting vs fine-tuning, preparing {prompt, completion} or chat datasets, running SFT jobs, validating generation failures, and deciding adapter-only vs merged full-model distribution.
mistral-hidden-params-ft
by Haruk1y
Run end-to-end fine-tuning for request_text-to-hidden-params JSON prediction with Ministral-3-3B in this repository. Use when rebuilding train/validation/test splits from train-only data, converting data to prompt-completion format, launching TRL SFT on Hugging Face Jobs, validating strict JSON output behavior, merging LoRA adapters into full models, and reporting job/model status with links.