tinyllms/qwen2.5-7b-instruct-grpo-loo-domain-knowledge

Qwen2.5-7B-Instruct SFT+GRPO — LOO Domain Knowledge (12 steps)

GRPO-trained from tinyllms/qwen2.5-7b-instruct-sft-loo-domain-knowledge (itself SFT'd from Qwen/Qwen2.5-7B-Instruct) using QLoRA (4-bit NF4 quantization + LoRA adapters).

This is a capped variant (max 12 steps) of the GRPO stage of a leave-one-out (LOO) experiment: the model is trained on Game24 and AIME trajectories, deliberately excluding domain knowledge (GPQA) data. The held-out domain is later used to measure cross-domain transfer.

Training Configuration

• Algorithm: GRPO (Group Relative Policy Optimization)
• Learning rate: 5e-6
• Batch size: 2 per device
• Group size: 2 (completions per prompt)
• Max steps: 12
• Max completion length: 2048
• Precision: bf16
• Gradient checkpointing: enabled

QLoRA

• Quantization: 4-bit NF4 with double quantization
• LoRA rank: 64
• LoRA alpha: 128
• LoRA dropout: 0.05
• Target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj

Reward Functions

• correctness_reward: task-specific answer verification (binary 0/1)
• format_reward: checks for proper <answer> tag formatting, penalises excessive length

Datasets

Trained on two datasets (domain knowledge held out):

Dataset	Domain
tinyllms/game24-trajectories	Game of 24 — arithmetic reasoning
tinyllms/aime-1983-2023-trajectories	AIME — competition math

Leave-One-Out Design

Domain	Role
Game24	Train
AIME	Train
Domain Knowledge (GPQA)	Held out

Transfer is measured by evaluating on GPQA Diamond.

Infrastructure

• GPU: 2x NVIDIA H100 80GB (DDP)
• Framework: TRL + Ray Train (DDP)
• Tracking: Weights & Biases (project: pocket-sheet-grpo)
• Ray Job ID: raysubmit_YbfK7VEEYMGjEp9h