SmolLM2-360M-base-gsm8k-boxed-sft

SmolLM2-360M — boxed GSM8K SFT, 6 epochs

Supervised fine-tune of HuggingFaceTB/SmolLM2-360M (plain base, not Instruct) on 978 verified teacher traces for GSM8K in the "boxed" recipe. Checkpoint at 6 epochs (global step 180) = 5,868 demonstration presentations.

Why this epoch count

One of a ladder of SFT checkpoints whose demonstration budgets are matched to the hybrid (unified-training) arms of the GSM8K sweep, so that a staged/sequential arm (SFT then RL) can be compared against a hybrid arm (demos streamed during RL) at an identical total demonstration budget.

Training

• Corpus: 978 traces, Qwen2.5-3B-Instruct teacher, shortest-correct-fitting-budget selection, every trace verified 1.0 under the training grader.
• Prompt (verl standard, zero-shot): {question} Let's think step by step and output the final answer within \boxed{}.
• Targets end with EOS so the student learns to terminate.
• lr 1e-5, batch 32, cosine decay over the run's own epoch count (each ladder rung is a SEPARATE run — an intermediate checkpoint of a longer run is not an equivalent model, because the LR schedule differs).
• Full-model SFT (no LoRA), bf16, 1 GPU (FSDP world size 1).

Intended use

Initialization for RL. Evaluate with a strict grader that reads the last \boxed{...} span (unparseable or absent box scores 0); the campaign's reference protocol is n=128 samples, temperature 0.6, top_p 0.95, max 1024 new tokens, on the full 1,236-question GSM8K test split.

Reference numbers for the shorter rungs (same protocol, strict scorer): 6 ep = .0683 pass@1 / .6693 pass@64; 9 ep = .0718 / .6934; 12 ep = .0756 / .6990.