Training
Table | |
|---|
| Method | DPO (TRL + Unsloth) |
| beta | 0.1 |
| LoRA rank / alpha | 32 / 32 |
| Effective batch | 128 (8 × grad_accum 16) |
| Max seq / prompt length | 1024 / 768 |
| Learning rate | 2e-4, linear decay |
| Epochs | 3 (2,757 steps) |
| Hardware | 1× RTX 3090, ~6h17m |
Evaluation
Held-out 2% split + diversity on 100 sampled prompts (4 responses × 4 temperatures).
Table with columns: Metric, SFT baseline, This model (DPO)| Metric | SFT baseline | This model (DPO) |
|---|
| Eval loss | — | 0.457 |
| Reward accuracy (held-out) | 0.50 (chance) | 0.72 |
| Reward margin | — | 1.65 |
| Diversity — EAD | 0.1173 | 0.1193 |
| Diversity — SBERT | 0.2263 | 0.2322 |
| Diversity — Vendi | 2.7327 |
Takeaways: the model learned the preference (reward accuracy 0.50 → 0.72) while preserving output diversity (no mode collapse — all diversity metrics flat vs the SFT baseline). Training shows mild overfitting (train reward accuracy ~0.85 vs eval 0.72), so 3 epochs is the right length.
Intended use & limitations
- Same scope as the SFT base: a structured ML-paper research assistant, not a general chatbot. Best used via the
PaperResearcher task API from the SFT stage.
- At 135M parameters the model is capacity-limited — it learns task shape and preference, not deep factual recall. DPO sharpens which response style is preferred; it does not add knowledge.
- The reward/eval accuracy measures agreement with the LLM judge that created the preference data, so it is not a fully independent quality signal.
Reproduce
See dpo/DPO_SmolLM135M (run_dpo.sh, experiments.md, LEARNINGS.md).
