wenhuahuo/chip-0.8b

Chip

Chip is the main NL2Hull ship-design decision model. It adapts a Kev-style typed decision interface to Chinese natural-language ship-form editing requests and predicts probability distributions over ordered editing decisions.

This release contains the LoRA adapter and the learned decision head for the 0.8-billion-parameter Chip model. The adapter is applied to Qwen/Qwen3.5-0.8B-Base; the head.pt file is required by the NL2Hull/Kev inference implementation.

Files

• adapter_model.safetensors: LoRA adapter weights
• adapter_config.json: PEFT adapter configuration
• head.pt: learned typed-decision head
• tokenizer.json, tokenizer_config.json, chat_template.jinja: tokenizer files
• training_config.json, training_metrics.json: training provenance

Intended use

Use Chip for typed probability prediction over the NL2Hull ship-design decision interface. The predicted decisions can be projected into ordered free-form deformation actions and evaluated with the shared benchmark in the NL2Hull repository.

Training

Chip was initialized from the released Kev-0.8B decision checkpoint and adapted on all 88,604 records in the SDD Dataset training split for two epochs using rank-16 LoRA adaptation, a learning rate of 2e-5, bfloat16 weights, and an effective batch size of eight.

Evaluation

On SDDBench, Chip achieved 95.90% question accuracy and 99.32% FFD exact match, with an NLL of 0.0951, an expected calibration error of 0.0032, and a Brier score of 0.0551. The reported protocol evaluates probability distributions and projected action sequences; it does not regress continuous magnitudes or spatial ranges.

Usage

Clone the NL2Hull repository (https://github.com/wenhuahuo/NL2Hull) and install its dependencies. Load the Qwen base model, the adapter, and head.pt through the project’s Kev-compatible inference code. The model repository stores the task-specific adapter and head; the base model remains a separate dependency.

Limitations

Chip is specialized for the fixed NL2Hull typed decision vocabulary and the SDD Dataset distribution. It is designed for research evaluation and decision support, with geometric execution and constraint checks performed by the NL2Hull engine.

License and attribution

The adapter is released for research use with the terms applicable to the base model and the NL2Hull project. Users should review the Qwen base-model license and the NL2Hull repository license before redistribution or commercial use.

Paper

Chip is described in NL2Hull: A Natural Language-Driven Constrained Ship Design Decision Framework.

Citation

@article{huo2026nl2hull,
  title         = {NL2Hull: A Natural Language-Driven Constrained Ship Design Decision Framework},
  author        = {Huo, Wenhua and Han, Fenglei and Zhao, Wangyuan and Wu, Jialin and Han, Jiayi},
  year          = {2026},
  eprint        = {2610.09896},
  archivePrefix = {arXiv},
  primaryClass  = {cs.AI},
  url           = {https://arxiv.org/abs/2610.09896}
}