Modelpublic

cstr/confucius4-r2t2-GGUF

sourceHugging Faceotherupdated 2d agoView on Hugging Face

0likes595downloads

Model Card

Confucius4-R2T2 — GGUF

GGUF conversions of `netease-youdao/Confucius4-R2T2` for [CrispASR](https://github.com/CrispStrobe/CrispASR) (--backend qwen3).

R2T2 ("Real Real-Time Transcription") is NetEase Youdao's streaming fine-tune of Qwen3-ASR-1.7B. It emits append-only text, committing only stable prefixes, with 80 ms to 2 s decoding chunks. The architecture is stock Qwen3-ASR (the configs are identical to Qwen/Qwen3-ASR-1.7B), except that R2T2 ties lm_head to the token embedding. CrispASR runs it offline like any Qwen3-ASR model, and also in its realtime WebSocket session with R2T2's own streaming recipe.

Any modifications made to the original model in this Derivative Work are not endorsed, warranted, or guaranteed by the original right-holder of the original model, and the original right-holder disclaims all liability related to this Derivative Work.

Files

File	Notes
`confucius4-r2t2-f16.gguf`	F16
`confucius4-r2t2-q8_0.gguf`	Q8_0
`confucius4-r2t2-q4_k.gguf`	Q4_K (default for `-m confucius4-r2t2`)
`MODEL_LICENSE`	the NetEase Youdao Model Use License Agreement, which applies to every file here

Converted from revision 185ce639 (SHA-256 of model.safetensors: cc4d5324…4610dc). All files carry qwen3asr.streaming_recipe = r2t2, which turns on the streaming session by default.

Streaming

The realtime endpoint (crispasr --server --ws-port N, then ws://…:N+1/v1/realtime) runs R2T2's prefix-rollback algorithm. It is a port of R2T2ASRModel.streaming_transcribe driven by the repository's example.py schedule: 160 ms steps, a 160 ms first-chunk lookahead, rollback of 1 token, and adaptive max_new_tokens. Deltas are append-only, following R2T2's own ws_server.py. Tunable with CRISPASR_QWEN3_STREAM_STEP_MS, CRISPASR_QWEN3_STREAM_LOOKAHEAD_MS and CRISPASR_QWEN3_STREAM_UNFIXED_TOKENS.

Verification

Against the upstream qwen_asr package (float32, CPU) on jfk.wav (English) and a Mandarin clip:

check	F16	Q8_0	Q4_K
encoder output (reference mel), worst-frame cos	1.000000	0.9993	0.9993
first-step logits cos	0.999998	0.9964	0.83
offline transcript vs upstream	identical (en, zh)	identical (en, zh)	same words, punctuation differs
streaming final text vs upstream `example.py`	identical (en, zh)	—	identical (zh), punctuation (en)

The streaming reference ran R2T2's own example.run_streaming unmodified, with its vLLM call served by the same weights through transformers. Call by call, CrispASR reproduces the same audio lengths and token budgets, and the same generated text on 66 of 68 (en) and 73 of 81 (zh) model calls. The rest are near-tied greedy steps that re-converge to the identical final transcript.

Licence

These files are a Derivative Work under the NetEase Youdao Model Use License Agreement (see MODEL_LICENSE), not an open-source licence. Obligations include the following:

—Retain the agreement with every copy.
—Pass these terms on to anyone you redistribute to.
—Make the disclaimer above when you distribute a derivative.
—Obtain a separate licence from NetEase if your products exceed 100 M monthly active users or RMB 1 B annual revenue.
—Keep the model out of the high-risk uses listed in §4.2.

The inference code in CrispASR is MIT and unaffected.