SPSD-RL-Qwen3-4B-Factory-MHTrue-2Ep-20260608

Qwen3-4B SPSD-RL LLaMA-Factory SFT, mask_history=true, 2 epochs

This model is a full-parameter fine-tune of Qwen/Qwen3-4B-Base on LorMolf/SPSD-RL using the repository LLaMA-Factory SFT pipeline.

This upload is intentionally separate from the earlier LorMolf/SPSD-RL model repo so previous uploaded models are not overwritten.

Training

• Base model: Qwen/Qwen3-4B-Base
• Dataset: LorMolf/SPSD-RL, data/*.jsonl
• Dataset revision: 76e62ee11f0b6b8e9a5511a7044a556f7c0c8e42
• Pipeline: src/training_eval/train_sft_factory.py
• Template: qwen
• Supervision: prompt/completion assistant-turn expansion
• LLaMA-Factory masking: train_on_prompt=false, mask_history=true
• Sequence length: 16384
• Epochs: 2
• Per-device train batch size: 1
• GPUs: 4
• Gradient accumulation steps: 16
• Effective train batch size: 64
• Learning rate: 2e-5
• Warmup ratio: 0.03
• Scheduler: linear
• Precision: bf16
• Packing: true, neat_packing=true

Final training metrics from the local run:

• train_loss: 0.06013819321350911
• train_runtime: 22:30:13.22
• train_steps_per_second: 0.011
• Final epoch: 2.0

W&B run: https://wandb.ai/lorenzo-molfetta/olmo-spiral-sft/runs/qr44c5qn

Notes

This run was launched before train-time validation was added to the factory pipeline, so it has no validation metrics. Use the repository generation evaluation pipeline for downstream SPSD-RL benchmark results.