slm-125m-legal-instruct

A 125M-parameter Llama-architecture small language model, instruction-tuned for grounded legal & financial question answering.

• Base: thesreedath/slm-125m-base (125M, pretrained on legal/financial + web text, 10 epochs)
• Fine-tuning: SFT on ~7,760 RAFT-style grounded Q&A pairs across 4 task types (grounded QA, summarization, extraction, rewriting), synthesized with Gemini and quality-filtered (LLM judge + embedding dedup).
• Chat format: <|system|> / <|user|> / <|assistant|>; terminate with <|eos|> (id 1).
• Context length: 1024 tokens. Answer-only loss (context + question -> answer).

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
tok = AutoTokenizer.from_pretrained("Ashish-Ranjan/slm-125m-legal-instruct")
model = AutoModelForCausalLM.from_pretrained("Ashish-Ranjan/slm-125m-legal-instruct")
sys = "You are a helpful legal and financial assistant. Answer using only the provided context."
prompt = "<|system|>" + sys + "<|user|>" + context + "\n\n" + question + "<|assistant|>"
ids = tok(prompt, return_tensors="pt").input_ids
out = model.generate(ids, max_new_tokens=200, do_sample=True, temperature=0.7, eos_token_id=1)
print(tok.decode(out[0][ids.shape[1]:], skip_special_tokens=True))