arc-agi2-qwen3-4b-exp3

Qwen3-4B-Thinking-2507 SFT on ARC-AGI-2 training data, third experiment.

Run details

• Date: 2026-05-14 → 2026-05-15
• Hardware: 3 nodes × 8× H100 80GB = 24 GPUs, ms-swift Megatron
• Base model: Qwen3-4B-Thinking-2507 with tokenizer surgery (vocab=16, untied embed/lm_head)
• Train iters: 7831
• gbs: 120
• max_length: 12288
• lr: 0.0001 (linear decay to 1e-07, warmup 200)
• adam_beta1/2: 0.9 / 0.98
• MFU: ~34% (10.7K tokens/sec/GPU)
• Checkpoint: checkpoint-7831

What's different from Exp 2

Exp 2 had a silent loss-mask bug: --template_backend jinja returned loss_scale=1.0 for the entire conversation, meaning the model was trained to predict the input grid in addition to the output grid. ~50% of gradient capacity was wasted.

Exp 3 fixes this with:

• Custom nvarc_qwen template registered via --external_plugins
• --template_backend swift --loss_scale default for proper role-based masking
• Parity verified 100/100 against NVARC's NeMo-RL reference on tokens + label positions

See feedback_msswift_jinja_loss_mask_bug.md in the project memory dir.

Loss summary

• First train_loss (iter 1): 0.2069
• Last train_loss (iter 7830): 0.0661
• Last eval_loss (iter 7831): 0.0961
• Total logged iters: 784

Files

• *.safetensors — bf16 weights
• config.json, tokenizer* — model + tokenizer (vocab=16)
• args.json — full ms-swift training args
• logging.jsonl — every 10-iter metric (loss, grad_norm, lr, mem, train_speed)
• runs/ — tensorboard event files

Citation

Built on NVARC's 1st-place 2025 ARC Prize solution; reproduction effort.