spider LLaVA-0.19 Qwen3.5-9B 128K — checkpoint-1700
Full-parameter behavior-cloning SFT snapshot.
- Base model:
Qwen/Qwen3.5-9B - Source run:
spider3_llava019_qwen35_128k_4n_20260911T120943Z - Snapshot:
checkpoint-1700 - Trainer progress: packed step 1700 (scheduler cap 10315)
- Approx game passes: 2.0238 (about 840 packed steps / pass at GBS=128, 128k)
- Spider passes=3.0 + LLaVA 0.19. TRAIN_ITERS=10315 is an unpacked-row cap; packed steps are much fewer.
- Training: TP=4, PP=1, 4 nodes, GBS=128, sequence packing 128K, MTP=1
- Special tokens:
<|action_start|>,<|action_end|>,<|action_sep|>,<|thought_start|>,<|thought_end|>