4fhct4sd/qwen3.5-4b-posture-checkpoints

Qwen3.5-4B Posture — adapters and checkpoints

The repository root contains the final phase 2 adapter and tokenizer, directly loadable using PEFT/Unsloth. Base model: Qwen/Qwen3.5-4B.

• phase1/final: scatter-phase adapter.
• phase1/checkpoint-95: resumable Trainer checkpoint, including optimizer/scheduler/RNG state.
• phase2/final: final crystal-phase adapter, continuing phase 1.
• phase2/checkpoint-42: resumable Trainer checkpoint.
• training-config.yaml: configuration with credentials removed.
• environment-versions.txt: training environment package versions.

Merged model: https://huggingface.co/4fhct4sd/qwen3.5-4b-posture

Resume from the checkpoint-N directories with the matching training script and datasets, not from final. Datasets are not uploaded. Transformers 5.5.0 is the locally tested inference version; avoid the observed Studio downgrade to 5.3.0 for this model. No benchmark evaluation has been run.