harshavaishnav/LMPlayschool_Submission

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:

playpen eval --suite all

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:

• 13hrs

Frameworks:

• Transformers
• PEFT
• TRL
• Unsloth

---

Evaluation command:

playpen eval --suite all

---

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.

---