This repository is the standalone BF16 inference checkpoint for the Voice Light conversational model. It merges the Voice Light tool-use LoRA into the original Qwen3-1.7B weights. It does not require PEFT or dynamic LoRA loading at inference time.

Sources

Base model: Qwen/Qwen3-1.7B at revision 70d244cc86ccca08cf5af4e1e306ecf908b1ad5e
LoRA adapter: BertilBraun/qwen3-1.7b-voice-light-tool-use-lora at revision 5c88c42fe092159e960126823223c3f46056cd7d
Training dataset: BertilBraun/voice-light-tool-use-synthetic

Refer to those repositories for the original model documentation, adapter training information, licenses, and intended-use constraints.

The dataset includes the exact teacher-led v22 generator snapshot, locked environment, tests, teacher-server command, end-to-end reproduction launcher, immutable teacher revision, generation manifests, deterministic scenario seeds, and file hashes under reproducibility/source-8c32f70.

Merge

Method: PEFT merge_and_unload(safe_merge=True)
Merge device: CPU
Weight format: safetensors with a 5GB maximum shard size
Weight dtype: BF16
Hugging Face Hub: 1.18.0
Safetensors: 0.8.0
Transformers: 5.14.1
PEFT: 0.18.1
PyTorch: 2.11.0+cu130

The exact machine-readable source revisions and merge settings are included in merge-provenance.json. The destination repository expected when this artifact was built is BertilBraun/qwen3-1.7b-voice-light-tool-use-merged.