| library_name: transformers.js | |
| base_model: Qwen/Qwen3.5-0.8B | |
| tags: [qwen, asmr, lora, onnx, webgpu] | |
| # WebASMR director (Qwen3.5-0.8B + LoRA, q4f16) | |
| Browser director for the WebASMR night studio. The `DownFlow/webasmr-qwen3.5-0.8b-lora` | |
| adapter was merged into the text-only base and requantized to 4-bit blockwise | |
| ONNX (`MatMulNBits`, block 32) without changing the hybrid-cache graph. | |
| - `onnx/model_q4f16.onnx` + `onnx/model_q4f16.onnx_data`: prefill/decode graph | |
| - tokenizer/config from `onnx-community/Qwen3.5-0.8B-Text-ONNX` | |
| - prompt format: raw `Scenario: … / Tone: … / Tags: …`, then `[tone:` | |
| - generation: temperature 0.8, top_p 0.9, repetition 1.1, stop at `<|im_end|>` | |
| The adapter is trained on a purchased ASMR corpus; use is non-commercial. | |