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

Qwen2.5-7B-Instruct SFT — LOO Domain Knowledge

Fine-tuned from Qwen/Qwen2.5-7B-Instruct using QLoRA (4-bit NF4 quantization + LoRA adapters, merged before upload).

This is the SFT 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

• Learning rate: 2e-5 (cosine schedule, 5% warmup)
• Batch size: 1 per device, gradient accumulation 16 (effective batch size 32 with 2 workers)
• Epochs: 1
• Max sequence length: 16384
• Precision: bf16
• Weight decay: 0.01

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

Loss

• completion_only_loss: prompt tokens are masked; loss is computed only on assistant completion tokens
• Dataset is converted from messages to prompt/completion format before training

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

Examples exceeding max_seq_len are filtered out. A 10% holdout is used for evaluation (eval runs every 10 steps).

Leave-One-Out Design

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

The GRPO stage follows using tinyllms/qwen2.5-7b-instruct-grpo-loo-domain-knowledge, trained on the same two datasets. Transfer is measured by evaluating on GPQA Diamond.

Infrastructure

• GPU: 2x NVIDIA H100 80GB (DDP)
• Framework: TRL 0.29 + Ray Train
• Tracking: Weights & Biases (project: pocket-sheet-sft)
• Ray Job ID: raysubmit_A55M5NnZckrXmfWN