| --- |
| license: other |
| pretty_name: RASST Demo ACL6060 Chinese Streaming Segments |
| tags: |
| - rasst |
| - sglang-omni |
| - qwen3-omni |
| - speech-translation |
| - streaming-translation |
| - acl6060 |
| - profiling |
| --- |
| |
| # RASST Demo ACL6060 Chinese Streaming Segments |
|
|
| This dataset contains the prepared ACL6060 en-to-zh streaming evaluation |
| segments used by the RASST demo and sglang-omni thinker decode profiling |
| workflow. |
|
|
| It is a convenience bundle for reproducing N=32 streaming load experiments. The |
| source ACL6060 data is public; this repository packages derived 16 kHz mono WAV |
| clips plus SimulEval-compatible source and target lists. |
|
|
| ## Contents |
|
|
| - `seg/`: 468 cropped WAV segments. |
| - `segments.source`: absolute segment paths from the original Taurus |
| preparation run, one per line. |
| - `segments.target`: Chinese references, one per line. |
| - `audio.yaml`: SimulEval-style segment manifest with `wav`, `offset`, and |
| `duration`. |
| - `ref.txt`: Chinese references for scoring. |
| - `source_text.txt`: English source text. |
| - `segments.meta.jsonl`: per-segment provenance. |
| - `SHA256SUMS`: checksums for all uploaded files except itself. |
| - `dataset_manifest.json`: compact generation and provenance metadata. |
|
|
| ## Dataset Summary |
|
|
| - Task: English audio to Chinese streaming speech translation. |
| - Segment count: 468. |
| - Talks: 5 ACL 2022 talks. |
| - Audio format: WAV, 16 kHz, mono. |
| - Total prepared audio duration: 3107.332 seconds. |
|
|
| ## Reproduction Usage |
|
|
| Download: |
|
|
| ```bash |
| huggingface-cli download gavinlaw/rasst-demo-acl6060-zh-segments \ |
| --repo-type dataset \ |
| --local-dir /path/to/acl6060_zh_segments |
| ``` |
|
|
| Use with the sglang-omni diagnostic package: |
|
|
| ```bash |
| cd /path/to/sglang-omni |
| export DATA_DIR=/path/to/acl6060_zh_segments |
| |
| ENABLE_MIXED_CHUNK= CHUNKED_PREFILL_SIZE=0 \ |
| bash benchmarks/diagnostics/thinker_decode_forward/scripts/p1_nsys_decode.sh |
| ``` |
|
|
| ## Generation |
|
|
| The prepared directory was generated from the RASST ACL zh main-result input |
| manifest and ACL6060 full-talk WAV files using: |
|
|
| ```bash |
| python eval/streaming_sst/prepare_acl6060_segments.py \ |
| --out-dir /mnt/taurus/data2/jiaxuanluo/rasst_eval/acl6060_zh_segments |
| ``` |
|
|
| Equivalent generator in the sglang-omni fork: |
|
|
| ```bash |
| python benchmarks/diagnostics/qwen3_omni_tp_vllm_gap/eval/prepare_acl6060_segments.py \ |
| --out-dir benchmarks/diagnostics/thinker_decode_forward/artifacts/data/acl6060_zh_segments |
| ``` |
|
|
| ## Provenance |
|
|
| Original local preparation path: |
|
|
| ```text |
| /mnt/taurus/data2/jiaxuanluo/rasst_eval/acl6060_zh_segments |
| ``` |
|
|
| RASST release data source: |
|
|
| ```text |
| gavinlaw/rasst-main-result-data |
| ``` |
|
|
| The upstream ACL6060 source data is public. Please check the upstream dataset |
| terms before redistributing modified copies outside this research reproduction |
| context. |
|
|
