ifx-pse-sys-ml/flame-27m-instruct

Fix generate(): implement no_repeat_ngram_size, add penalize_prompt; correct the generation docs

Card: document generation parameters (repetition_penalty, no_repeat_ngram_size, temperature, top_p, max_new_tokens) + overrides

Set default decoding: repetition_penalty 1.3 + no_repeat_ngram 3 (curbs multi-sentence repetition)

Update to Tier-A SFT (richer instruction mixture; better instruction-following)

Add flame-27m (300B-token base + SmolTalk instruct)

initial commit