Qwen3.5-9B-decision-subtrajectory-32k-epoch2

Qwen3.5-9B decision subtrajectory SFT — epoch 2

BF16 causal language model, full SFT at step 9312. Training: 37,242 decision-boundary samples, maximum sequence length 32768, precomputed ownership labels (historical assistant tokens masked), no truncation or split packing. Validation split excluded from training. Learning rate 1e-5 constant, global batch 8, 8 B200, FSDP2, 26 layers activation checkpointing, chunked NLL 1024. No downstream quality evaluation has been performed.

Export includes tokenizer and training chat template. Original resumable optimizer/checkpoint state and full logs remain on NAS. This repository contains inference model weights, not optimizer state.