mllm-mmr1-ttrl-gemma3-4b-full · best checkpoint (step 540)

• Training data: MMR1/MMR1-Math-RL-Data-v0 (5,782 prompts after cap+prune)
• Budget: exactly 1 epoch = 722 steps, effective batch 64 (bs 1 x grad_accum 8 x 8 GPUs), lr 1e-6, num_generations 8, max_completion_length 1024, seed 42. --max_steps was never set; the step count falls out of the dataset size (5,782 prompts / 8 prompts per step).
• Eval: MathVista testmini-150 (in-loop, 152 rows after distributed padding), every 20 steps, plus a step-0 baseline
• This checkpoint: step 540 of 722 — the best one
• Its score: 0.3750

Eval curve

Files for audit

• train.log — the complete training log this checkpoint came from
• eval_curve.csv — the table above, machine-readable
• run_config.json — config as the trainer saw it