SmolLM2-360M-instruct-gsm8k-boxed-sft

SmolLM2-360M-instruct — GSM8K boxed-answer SFT

HuggingFaceTB/SmolLM2-360M-Instruct fine-tuned for 6 epochs on 978 verified teacher traces (Qwen2.5-3B-Instruct, K=8 samples/question, kept only if correct under a strict boxed grader, finish_reason=stop, and ≤511 tokens under this model's tokenizer; shortest kept trace per question).

Prompt format

One user message, zero-shot:

{question} Let's think step by step and output the final answer within \\boxed{}.

Template applied at training and inference: the model's own chat template (system + user turns). Targets end with this tokenizer's EOS.

Training

Full-model SFT (no LoRA), lr 1e-5, batch 32, 6 epochs (183 steps), loss on the response tokens + terminal only. verl sft_trainer.

Evaluation (GSM8K test, 1,236 questions, n=128, temp 0.6, top_p 0.95)

Strict = Math-Verify applied to the last \boxed{} span only, no fallback. Under a lenient whole-response grader the numbers are within .003 — the model boxes what it knows.

Provenance

Trained 2026-08-25 on the unified_train branch of an off-policy MaxRL research codebase (SmolLM2/GSM8K hybrid-SFT campaign). SFT questions are disjoint from the campaign's RL prompt pool and test set (verified exact and at skeleton level). Intended as an RL initialization / staged-training control.