Views
No views yet
model.safetensors
holds TurboQuant-quantized weights in helix's packed layout, with the
quantization index embedded in the safetensors header metadata. It loads with
the bee/helix engine, not AutoModel.| tower | format | notes |
|---|---|---|
| text | tq4_1s | 4-bit TurboQuant blocks (32-value WHT block, dual fp16 scales) |
| audio | tq4_1s | 4-bit TurboQuant blocks |
Qwen/Qwen3-ASR-ForcedAligner-0.6B weights with
hx packer. Boundary accuracy was validated against a tq8-text oracle on a
116-clip cut corpus: every clip agrees with the tq8 reference to within one
80 ms time bucket (zero multi-bucket errors), at ~34% smaller on disk
(~558 MiB vs ~845 MiB).model.safetensors — packed TQ weights + embedded turboquant.indexconfig.json — aligner config (classify_num, timestamp_segment_time,
token ids, text hidden_size)tokenizer.json, tokenizer_config.json, vocab.json, merges.txt — Qwen3 tokenizerchat_template.json, generation_config.json, preprocessor_config.json — aux