4fhct4sd/qwen3.5-4b-posture

Qwen3.5-4B Posture — merged model

Final phase 2 LoRA adapter merged into Qwen/Qwen3.5-4B, exported as 16-bit safetensors. The adapter continues phase 1 (scatter) with phase 2 (crystal) training.

Adapters and resumable checkpoints: https://huggingface.co/4fhct4sd/qwen3.5-4b-posture-checkpoints

Use Transformers 5.5.0 for the tested inference runtime. The local Studio 5.3.0 sidecar produced invalid-probability CUDA assertions during multi-turn chat; 5.5.0 passed a five-turn test. This is limited validation, not a general quality or stability guarantee. The model may invent conversational turns; generated claims about itself are not verified facts.

No evaluation benchmark has been run. Training data is not included in this repository.