slm-125m-qa

125M legal SLM, closed-book QA fine-tune of thesreedath/slm-125m-base. Answers questions from its own weights (no context passage), and refuses when it does not know. Trained on ~7.1k self-contained, faithfulness-judged, decontaminated Q&A pairs (10% refusals). Loss on answer tokens only.

Chat template: <|bos|><|system|>\n{SYSTEM}<|eos|>\n<|user|>\n{QUESTION}<|eos|>\n<|assistant|>\n