Qwen3.5-4B LoRA SFT for the LM Playschool Challenge
A LoRA SFT fine-tune of the base Qwen/Qwen3.5-4B model on the dataset of harshavaishnav/Dialogue_games, submitted to the LM Playschool Challenge.
- Model:
harshavaishnav/LMPlayschool_Submission
- Base Model:
Qwen/Qwen3.5-4B
- Method: LoRA Supervised Fine-Tuning (SFT)
Headline Results
| Model | clemscore | statscore |
|---|
| Qwen/Qwen3.5-4B (baseline) | 37.33 | 52.07 |
| This model | 38.03 | 48.07 |
| Δ | +0.7 | -4 |
Evaluated using:
Per-game evaluation results are available in the qwen_val.json file , clem and static directories.
Training Methodology
This model was fine-tuned using Supervised Fine-Tuning (SFT) with LoRA adapters on the success-filtered Playpen preference dataset.
Main characteristics:
- Base model: Qwen/Qwen3.5-4B
- LoRA fine-tuning
- Supervised Fine-Tuning (SFT)
- Instruction-response pair training
- One training epoch
- Gradient checkpointing enabled
- Mixed precision (bf16)
- Optimizer: AdamW
Data Usage
Dataset:
- harshavaishnav/Dialogue_games
Tokenization:
- Qwen chat template
- Maximum sequence length: 2048
No external datasets were used.
Hyperparameters
| Hyperparameter | Value |
|---|
| Base Model | Qwen/Qwen3.5-4B |
| Training Method | LoRA + SFT |
| LoRA Rank (r) | 32 |
| LoRA Alpha | 64 |
| LoRA Dropout | 0.05 |
| Learning Rate | 2e-4 |
| LR Scheduler | Cosine |
| Warmup Ratio | 0.03 |
| Per Device Batch Size | 2 |
| Gradient Accumulation Steps | 8 |
| Effective Batch Size | 16 |
| Max Sequence Length | 2048 |
| Training Epochs | 1 |
| Precision | bf16 |
| Gradient Checkpointing | Enabled |
| Optimizer | AdamW |
Compute
Training Hardware:
- 1 NVIDIA RTX 5000 Ada Generation GPU
Training Time:
Frameworks:
- Transformers
- PEFT
- TRL
- Unsloth
Evaluation command:
Design Decisions
The following design choices were made during training:
- LoRA-based Supervised Fine-Tuning (SFT) instead of full-parameter fine-tuning.
- Sequence length set to 2048 to balance context length and computational efficiency.
- Gradient checkpointing enabled for improved memory efficiency.
- Cosine learning-rate scheduler with warmup.
- bf16 mixed precision training.
- AdamW optimizer used for stable optimization.
Limitations
- Performance depends on the quality and diversity of the supervised training dataset and may not generalize to unrelated tasks.
- The model may still produce incorrect, incomplete, or hallucinated responses for unseen or complex tasks.
- Long-context reasoning beyond the training sequence length may be limited.
- The model may inherit biases from both the base model and the supervised training data.
License
This model inherits the Apache-2.0 license from the Qwen base model.
