zhiyuanhucs/genshin-llava019-qwen3.5-9b-128k
079
genshin LLaVA-0.19 Qwen3.5-9B 128K — checkpoint-2422
Full-parameter behavior-cloning SFT snapshot.
- Base model:
Qwen/Qwen3.5-9B - Source run:
genshin3_llava019_qwen35_128k_4n_20260902T181636Z - Snapshot:
checkpoint-2422 - Trainer progress: packed step 2422 (scheduler cap 34604)
- Approx game passes: 2.8833 (about 840 packed steps / pass at GBS=128, 128k)
- Genshin passes=3.0 + LLaVA 0.19. One Genshin pass is ~840 packed steps (GBS=128, 128k), not 34604/3.
- 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|>
