Qwen-GuidePlay-2B

Public validation result:

• Clemscore: 57.12
• Statscore: 42.68

Method

The model was trained with a three-stage interaction-learning pipeline:

1. Success SFT on complete successful Playpen dialogues.
2. Weighted turn-level SFT on state-action examples extracted from successful trajectories.
3. Lightweight LLM-guided augmentation, where a stronger teacher model (Gemma-4-31B) was used to judge selected examples and generate format-repair examples.

The final submitted model used the following best mixture:

• 30,000 original Stage 2 weighted-turn rows
• 1,000 LLM-judged/reweighted rows
• 100 LLM-generated format-repair rows

The teacher model did not create new gold actions. Gold actions always came from successful Playpen traces. The teacher only scored examples and generated realistic bad outputs for repair training.

Data Usage Breakdown

The model used four data components:

1. Successful Playpen dialogue trajectories for Stage 1 success SFT.
2. Turn-level state-action examples extracted from successful trajectories for Stage 2 weighted SFT.
3. LLM-judged Stage 2 rows, where a stronger teacher model scored selected examples.
4. LLM-generated format-repair rows, where realistic invalid outputs were generated and the model was trained to produce the clean gold action.

No private evaluation data was used for training.

Evaluation

Model / Method	Clemscore	Statscore
Base Qwen3.5-2B	13.05	44.02
Stage 1 Success SFT	45.76	39.74
Stage 2 Uniform	46.18	40.70
Stage 2 Weighted-turn 0.25	53.32	41.61
Qwen-GuidePlay-2B	57.12	42.68

Training Hyperparameters

The final submitted model was produced using the following training setup for the final model:

Setting	Value
Base model	Qwen/Qwen3.5-2B
Starting checkpoint	Stage 1 success SFT adapter
Training objective	Weighted supervised fine-tuning
Max sequence length	1024
Epochs	0.25
Learning rate	5e-5
Per-device train batch size	1
Gradient accumulation steps	8
Precision	bf16
Final format	Full merged model checkpoint

Earlier stages were used to construct the training pipeline: Stage 1 trained on complete successful dialogues, and Stage 2 trained on weighted turn-level state-action examples. The final submitted checkpoint was produced by merging the resulting adapter into the base model.

Design Decisions and Reproducibility

The strongest non-LLM baseline was weighted turn-level SFT, so the final model kept Stage 2 weighted-turn training as the main backbone. LLM-guided data was added only lightly to avoid over-specialisation.

The teacher model did not create new gold actions. The correct action always came from the original successful Playpen trace. The teacher only provided quality scores and generated realistic bad-output variants for repair training.

For Playpen evaluation, the model was registered with left padding and eos_to_cull set to <\\|im_end\\|>. The raw tokenizer EOS token is <|im_end|>; the escaped value is used by the evaluator for output post-processing.

Compute

Training was performed on one 80 GB NVIDIA A100 GPU using LoRA-style adapter training followed by merging into the base model.

LM Playschool Evaluation Registry Entry

For LM Playschool / Playpen evaluation, the model was registered with the following local Hugging Face backend configuration:

{
  "model_name": "Qwen-GuidePlay-2B",
  "backend": "huggingface_local",
  "huggingface_id": "syedhuq/qwen-guideplay-2b",
  "release_date": "2026-07-06",
  "open_weight": true,
  "parameters": "2B",
  "languages": ["en"],
  "context_size": "262k",
  "license": {
    "name": "Apache-2.0",
    "url": "https://www.apache.org/licenses/LICENSE-2.0"
  },
  "model_config": {
    "requires_api_key": false,
    "premade_chat_template": true,
    "eos_to_cull": "<\\|im_end\\|>",
    "padding_side": "left"
  }
}