machine5_continue_stable_expand13000_20260520_lr2e-7_e1

This is a full-parameter fine-tune of Qwen3-VL-8B-Thinking for visual symbolic regression: given a rendered function plot, immediately call submit_expression with a compact executable NumPy expression.

Training Recipe

• Parent checkpoint: machine5_rendered_stable_medium6000_20260520_lr5e-7_e1
• Training data: 13000 train-only rendered image/function pairs
• Source mix: official 7700, poly 3900, closure 1400
• Difficulty mix: easy 3073, medium 8129, hard 1198, expert 300, extreme 300
• Teacher trace: none
• Target format: compact submit_expression tool call with true expression
• Prompt/reasoning: image-only direct tool-call prompt, no reasoning
• Key protocol fix: stripped empty Qwen3 <think> template during SFT
• LR: 2e-7
• Global batch: 8
• Steps: 1625 optimizer steps, 1 epoch

No dev/test answer trajectories were used as training data.

Evaluation

Evaluated with 8 vLLM services, 60 workers, max_tokens=16000, thinking disabled, and direct tool-call extraction.

Balanced60:

• acc@0.99=0.35
• acc@0.95=0.3833
• acc@0.9=0.4
• acc@0.8=0.45
• null=0/60
• finish_reason={"stop": 60}
• mean latency: 6.276s

Official dev 300:

• acc@0.99=0.28
• acc@0.95=0.2967
• acc@0.9=0.31
• acc@0.8=0.34
• null=0/300
• finish_reason={"stop": 300}
• mean latency: 4.411s

Intended Use

This checkpoint is an experimental model for visual symbolic regression research. It is tuned for concise tool-call outputs and may be over-specialized to rendered single-variable function plots.