slm-125m-instr

125M legal SLM, instruction-tuned on top of thesreedath/slm-125m-qa (lineage: base -> closed-book QA SFT -> instruction SFT). Trained on ~6.5k domain-grounded synthetic instructions (summarize / extract / rewrite / classify / explain / draft / format-constrained tasks), every example compliance- and groundedness-judged. Loss on response tokens only.

Prompts may carry the working text inline after the instruction: <instruction>\n\nTEXT:\n<excerpt>.