---
license: mit
task_categories:
- automatic-speech-recognition
language:
- en
pretty_name: agwer ASR batch test v0
size_categories:
- n<1K
---

# agwer_asr_batch_test_v0

Example batch inputs for the [agwer](https://github.com/huckiyang/agwer)
CLI: 30 real ASR error-correction sessions (Whisper 5-best decodes from the
MIT-licensed HyPoradise benchmark, WSJ, with the verbatim outputs of a real
LLM corrector), wrapped in every input format agwer accepts:

| file | format | flag |
|---|---|---|
| `input.jsonl` | native agwer records | (default) |
| `input.openai.jsonl` | OpenAI chat sessions (`messages`) | `--format openai` |
| `input.anthropic.jsonl` | Anthropic Messages + structured output | `--format anthropic` |
| `input.sharegpt.jsonl` | ShareGPT conversations | `--format sharegpt` |
| `input.parquet` | columnar batch | `--format parquet` |

All five carry the same 30 sessions, so every format reports identical
metrics. `--format auto` (the default) detects each of them.

```bash
pip install "agwer[parquet]"
hf download huckiyang/agwer_asr_batch_test_v0 --repo-type dataset --local-dir batch
agwer batch/input.openai.jsonl
agwer batch/input.parquet --json
```

Expected report: WER 1-best 6.4%, corrected 3.1%, o_nb 4.8%, o_cp 1.7%,
RIR (rho) 2.000, HER 0.286.

In the chat formats, the corrected transcript is extracted from the last
assistant turn (structured outputs with a `"corrected"` key are unwrapped);
`reference` and `nbest` ride along as top-level keys.