Jakimovski (2025) —
tiny-reasoning-language-model,
with code at github.com/Shekswess/tiny-reasoning-language-model. Not a
paper; the second source in this record filed with a url: alone.
Key takeaways
- Base: SmolLM2-135M-Instruct. Three stages of full fine-tuning on one AMD
MI300X: SFT on 58k dialogues without chain of thought (3 epochs, peak LR
3e-4, cosine); SFT on 78k reasoning traces wrapped in
<think>tags added as special tokens (1 epoch, LR 3e-4); DPO on 50k chosen/rejected pairs (1 epoch, LR 1e-5, β 0.1). - The DPO stage converges at this scale and improves reasoning style and preference adherence — the fact LIT-119 cites when asking whether DPO would work at 90M (SOTA-126).
- Against the base model: ARC-Challenge 40.6 vs 37.3, BBH 36.8 vs 28.2, MMLU 35.0 vs 29.3, IFEval 35.5 vs 29.9, GSM8K 2.6 vs 1.4.
- The author’s own caveat: a research prototype, not for use.
Standing in the anthology
One run, no ablations, no baseline other than the untouched base model; here because LIT-119 leans on it as the prior for DPO at 100M, and because it is the second non-paper source the record carries (ADR-009). Its recipe is a curriculum — pretrain, SFT, then DPO — which is the approach LIT-119’s anti-curriculum result argues against for the pretraining half.