We present Qwen-GuidePlay-2B, a 2B-parameter language model for dialogue-game interaction. We fine-tune Qwen3.5-2B using three steps: a) SFT on only successful game trajectories from Playpen, b) weighted turn-level SFT, and c) teacher-guided SFT. The teacher model (which is a larger model) is only used to fix formatting and evaluate examples, but does not create new gold actions. Our final model scores 57.12 clemscore and 42.68 statscore on the public Playpen validation. In the officially released challenge results, our model obtains the second-highest Playpen clemscore delta among submitted systems (which is approximately +36 over its base model). Our findings suggest that imitating full trajectories helps with playability, while turn-level and teacher-guided training usually improve decision-making and increase the overall score. Alternative procedurally heavy approaches like replay-repair and hard-example mining did not help, which suggests that small models are performant simply by using careful curation strategies rather than aggressive changes. We make available both the model and the code for reproducibility.
Figures & tables
Component
Count
Share
Original Stage 2 weighted-turn rows
30,000
96.5%
Teacher-judged/reweighted rows
1,000
3.2%
Teacher-generated repair rows
100
0.3%
Total
31,100
100%
Table 1: Final Stage 3 training mixture.
Setting
Stage 1
Stage 2
Final
Initialisation
Base + new LoRA
Stage 1 adapter
Stage 1 adapter
Data
Success dialogues
∼ 106k turn rows
31,100-row mix
Loss tokens
All
Completion-only
Completion-only
Example weights
None
Heuristic
Heur./teacher/repair
Learning rate
2e-4
5e-5
5e-5
Epochs
1.0
0.25
0.25
Table 2: Per-stage training configuration.
Model / Method
Avg % Played
Avg Quality
Clemscore
Statscore
Base Qwen3.5-2B official
-
-
13.05
44.02
Base Qwen3.5-2B local
28.57
41.70
11.91
-
Stage 1: Success SFT
85.60
53.46
45.76
39.74
Stage 2: Uniform weights
-
-
46.18
40.70
Stage 2: Weighted-turn
83.77
63.65
53.32
41.61
Stage 2: Weighted-turn (0.5 epoch)
82.34
64.45
53.07
40.62
Table 3: Main results across training stages ( – indicates that the corresponding metric was not recorded.)
Comparison
Δ Avg % Played
Δ Avg Quality
Base local vs. Stage 1
+57.03
+11.76
Stage 1 vs. Stage 2 Weighted
-1.83
+10.19
Stage 2 Weighted vs. Final
-0.47
+4.92
Table 4: Playability and quality differences across training configurations.
Method
Clemscore
Statscore
Interpretation
Stage 2 Weighted-turn
53.32
41.61
Strong non-teacher baseline
Stage 2 Weighted-turn, 0.50 epoch
53.07
40.62
More training did not help
RR
54.86
38.66
Replay/repair improved clem but hurt stat
RR-scaled merge, α=0.75
53.12
-
Scaling RR delta did not recover the gain
Outcome-based weighting
51.06
40.84
Outcome-based weighting was noisy
HEM
51.41
40.60
Loss-based hard mining underperformed
Table 5: Ablation results. Higher is better for both Clemscore and Statscore; α denotes the scaling applied to the replay–repair LoRA delta. A dash indicates that the corresponding metric was not recorded.
Appendix figures & tables1 asset
Supplementary material from the paper’s appendix.
Appendix
Placeholder
Replaced with
{context}
The message history before the assistant action, serialised as JSON and truncated to its final 7,000 characters (the most recent game state matters most for the next action).
{target}
The original gold assistant answer from the Stage 2 row.
{original_weight}
The row’s existing Stage 2 heuristic value.
{row_index}
The row’s index in the Stage 2 file, used for resumable processing and provenance tracking.
{max_bad_outputs}
4 in all our runs.
Appendix
Table 6: Template placeholders in the judging prompt.