CLI

agwer results.jsonl
agwer results.jsonl --json --her-granularity token
agwer results.jsonl --raw          # skip default normalization

Each JSONL line is one utterance:

{"reference": "the cat sat", "corrected": "the cat sat",
 "nbest": ["the cat sad", "the cat sat"]}

nbest[0] must be the 1-best. "onebest": "..." may replace nbest (RIR and the oracles then unavailable; HER + WERs still reported).

The human-readable report rounds error rates to the standard ASR convention, one decimal (xx.x%). --json always emits full-precision floats, and the Python API never rounds; round only at the reporting step.

LLM output formats

Corrector outputs usually live inside chat logs. --format accepts them directly (auto, the default, detects each shape); the corrected transcript is taken from the last assistant turn, and structured outputs with a "corrected" key are unwrapped. reference and nbest (or onebest) always ride along as top-level keys.

OpenAI chat sessions (--format openai):

{"reference": "a b c", "nbest": ["a x c", "a b c"],
 "messages": [{"role": "system", "content": "correct the asr"},
              {"role": "user", "content": "hypothesis 1: a x c ..."},
              {"role": "assistant", "content": "a b c"}]}

Anthropic Messages with structured output (--format anthropic):

{"reference": "a b c", "nbest": ["a x c", "a b c"],
 "response": {"role": "assistant", "stop_reason": "end_turn",
              "content": [{"type": "text",
                           "text": "{\"corrected\": \"a b c\"}"}]}}

ShareGPT conversations (--format sharegpt):

{"reference": "a b c", "nbest": ["a x c", "a b c"],
 "conversations": [{"from": "human", "value": "hypothesis 1: a x c ..."},
                   {"from": "gpt", "value": "a b c"}]}

Parquet batches (--format parquet, columns reference, corrected, nbest) for efficient many-entry processing; needs pip install "agwer[parquet]".

A ready-made test set with the same 30 real correction sessions in all five formats lives at huckiyang/agwer_asr_batch_test_v0:

hf download huckiyang/agwer_asr_batch_test_v0 --repo-type dataset --local-dir batch
agwer batch/input.openai.jsonl      # identical report from every file

Benchmark your install:

python -m agwer.bench