R4 — Corrective feedback (DPO)
Part of a five-regime developmental sweep of post-training methods for dialogue-game competence (LM Playschool Challenge 2026, team DAIR).
A second DPO round on the merged R2 model using 201 first-move preference pairs from two rollout passes (greedy and t=0.7): 107 on-policy pairs from instances where the two passes disagreed in outcome (chosen = the model's own success, rejected = its own failure on the identical instance) and 94 hybrid pairs from instances failed in both passes (chosen = a stronger model's success on that instance — a recast). Same hyperparameters as R2.
Effect: 67.39 -> 67.64, the nominal best of the family, though the margin is within run-to-run variability. The on-policy pairs come from the model's competence frontier (instances of variable outcome).
All numbers are clemscore / statscore on the playpen validation split,
measured in a single frozen environment (Python 3.11, clemcore pinned via
playpen, clembench pinned requirements) with two upstream fixes applied:
a division-by-zero guard in the privateshared Game Master and the
punkt_tab NLTK resource for the IFEval scorer. Earlier revisions of this
card reported numbers from an unpinned environment; see the paper for the
environment-sensitivity analysis.
Checkpoint family (LM Playschool challenge, team DAIR)
| Regime | Repo | clem | stat |
|---|---|---|---|
| R1 imitation (SFT) | lm-playschool-qwen3.5-2b-sft | 55.61 | 43.87 |
| R2 outcome contrast (DPO) | lm-playschool-qwen3.5-2b-sft-dpo | 67.39 | 44.72 |
| R3 self-imitation (SFT) | lm-playschool-qwen3.5-2b-iter3 | 61.06 | 44.01 |
| R4 corrective feedback (DPO) | lm-playschool-qwen3.5-2b-iter4 | 67.64 | 44.31 |
| R5 GRPO (control) | lm-playschool-qwen3.5-2b-grpo-base-s42 | 62.43 | 44.19 |
| R5 GRPO + RND | lm-playschool-qwen3.5-2b-grpo-rnd-s42 | 67.44 | 43.53 |
Base model: Qwen3.5-2B (13.63 / 44.22 in the same environment). Paper: Raising a Small Language Model: From Imitation to Curiosity in Dialogue Games (LM Playschool Challenge 2026).