Skip to content

v0.4.2

Latest

Choose a tag to compare

@Liuziyu77 Liuziyu77 released this 15 Nov 01:27
· 1 commit to main since this release

Fixed

  • merge_checkpoint.py exports a LoRA fine-tune of a Valen-Decision-1.5 model: the exported config carries valen_full_attention_mask, valen_grounding and valen_decision_version, the trained grounding head is saved with the weights, and export validation runs shared_state for the block masks (question_blocks, bbox_blocks). The export manifest records the versions of the installed packages, so exporting with --attn-implementation sdpa works without flash-attn (#49).
  • merge_checkpoint.py runs directly on the output of python -m valen.train and a downloaded base model: it reads the final training metric from metrics.jsonl, and identifies a base model that ships no manifest by the hashes of its weight files. The exported usage note states the training step of the checkpoint (#49).

Code

  • The fine-tuning recipes sft_finetune_lora.json and sft_finetune_full.json start from models/Valen-Decision-1.5-0.8B, and the VisualDecisionBench evaluation script defaults to Valen-Team/Valen-Decision-1.5-2B, matching the README.