scripts/MODAL.md

# Running Modal ASR Baselines

BES-4805 evaluates NVIDIA Parakeet TDT 0.6B v3 as the batch baseline
`modal_nvidia_parakeet_tdt_0_6b_v3`. The checked-in deployment source is
`modal/parakeet_tdt_0_6b_v3.py`; do not use owner-private deployment code or
the old non-ticket model ID.

## Deployment Constants

- Benchmark model ID: `modal_nvidia_parakeet_tdt_0_6b_v3`
- Upstream model: `nvidia/parakeet-tdt-0.6b-v3`
- Upstream revision: `b51b7dc0fbf7f266a97880fb4b626c56d28f4b96`
- Modal app: `voice-code-bench-parakeet`
- Modal GPU: `L40S`
- Modal Volume: `voice-code-bench-parakeet-hf-cache`
- Runtime: Transformers `5.12.0`
- Endpoint secret: `MODAL_PARAKEET_ENDPOINT`

The app stores official Hugging Face weights in the Modal Volume at `/models`
and loads them once per warm container with `@modal.enter`. The benchmark sends
one complete 16 kHz mono WAV per request. There is no chunking, no prompts, no
custom vocabulary, and no post-ASR correction.

The Meta OmniASR section below covers BES-4806. Inworld STT is not part of these Modal deployment paths.

## Modal Credentials

Install and configure the Modal CLI/API credentials for the workspace that will
own the deployment:

```bash
python -m pip install modal==1.4.1
modal setup
```

Create a Modal proxy token for the environment that will serve the Web Function:

```bash
modal workspace proxy-tokens create
```

If your workspace uses scoped proxy tokens, allow the token in the deployment
environment. Store the endpoint and proxy token in the ignored `scripts/.secret`
file:

```dotenv
MODAL_PARAKEET_ENDPOINT=https://<your-parakeet-web-function>.modal.run
MODAL_PROXY_TOKEN_ID=wk-...
MODAL_PROXY_TOKEN_SECRET=ws-...
```

The entity-verification stage of `vcb-run` also needs `OPENAI_API_KEY=...`.
Never commit `scripts/.secret`.

## Warm The Model Cache

Run the cache warm-up before deployment:

```bash
modal run modal/parakeet_tdt_0_6b_v3.py::download_model
```

Verify in Modal logs that the download used
`nvidia/parakeet-tdt-0.6b-v3` at
`b51b7dc0fbf7f266a97880fb4b626c56d28f4b96` and wrote under
`/models/nvidia/parakeet-tdt-0.6b-v3`. The revision literal in
`modal/parakeet_tdt_0_6b_v3.py`, the volume directory, and the Hugging Face
revision should match before releasing results.

## Deploy

Deploy the Web Function:

```bash
modal deploy modal/parakeet_tdt_0_6b_v3.py
```

Copy the authenticated Web Function URL printed by Modal, or find it in the
Modal dashboard for `voice-code-bench-parakeet`, and save it as
`MODAL_PARAKEET_ENDPOINT`. The benchmark adapter accepts the exact Web Function
URL and normalizes trailing slashes.

The endpoint accepts JSON only:

```json
{
  "audio_base64": "<base64-encoded 16 kHz mono WAV bytes>",
  "audio_format": "wav"
}
```

It also accepts optional `"language": "en"` for manual compatibility checks,
but the registered benchmark configuration does not send a language hint. The
response body is exactly:

```json
{
  "transcript": "<verbatim model output>"
}
```

Requests must include the proxy-token headers:

```text
Modal-Key: <MODAL_PROXY_TOKEN_ID>
Modal-Secret: <MODAL_PROXY_TOKEN_SECRET>
```

The benchmark stores only the returned transcript verbatim.

## Longest-Recording Smoke Test

Create a one-row metadata file for the longest recording,
`audio_id=education_workplace_006` (`data/audio/261.wav`, 122.875 seconds):

```bash
mkdir -p runs/parakeet-longest-smoke
python - <<'PY'
import json
from pathlib import Path
source = Path('data/metadata.jsonl')
out = Path('runs/parakeet-longest-smoke/metadata.longest.jsonl')
for line in source.read_text().splitlines():
    row = json.loads(line)
    if row['audio_id'] == 'education_workplace_006':
        out.write_text(json.dumps(row, separators=(',', ':')) + '\n')
        break
else:
    raise SystemExit('education_workplace_006 not found')
PY
```

Run the smoke transcription without using `--limit 1`:

```bash
vcb-transcribe \
  --metadata runs/parakeet-longest-smoke/metadata.longest.jsonl \
  --stt-model-ids modal_nvidia_parakeet_tdt_0_6b_v3 \
  --output-dir runs/parakeet-longest-smoke
```

Inspect `runs/parakeet-longest-smoke/modal_nvidia_parakeet_tdt_0_6b_v3/transcripts.json`
and confirm it has one non-empty `model_transcript` for
`education_workplace_006`. Modal logs should show one complete request for the
122.875-second WAV and no client-side chunking.

## Full Run

With `MODAL_PARAKEET_ENDPOINT`, `MODAL_PROXY_TOKEN_ID`,
`MODAL_PROXY_TOKEN_SECRET`, and `OPENAI_API_KEY` configured, run:

```bash
vcb-run \
  --stt-model-ids modal_nvidia_parakeet_tdt_0_6b_v3 \
  --resume \
  --output-dir runs/parakeet-full
```

If interrupted, rerun the same command with `--resume`. After the run completes,
validate that transcripts and entity matches cover all 300 audio IDs in
`data/metadata.jsonl` order, then copy the generated entity-match artifact to
`baselines/predictions/modal_nvidia_parakeet_tdt_0_6b_v3.json` without editing
transcripts, entity decisions, or run metadata by hand.

Regenerate current aggregate outputs after the prediction artifact exists:

```bash
vcb-score-entities \
  --dataset-root . \
  --metadata data/metadata.jsonl \
  --entity-matches-dir baselines/predictions \
  --output-dir baselines \
  --scores-dir baselines/scores/entities \
  --output-csv baselines/results.csv

vcb-score-wer \
  --dataset-root . \
  --metadata data/metadata.jsonl \
  --transcripts-dir baselines/predictions \
  --output-dir baselines \
  --scores-dir baselines/scores/wer \
  --output-csv baselines/results.csv

vcb-make-figures --model-set all
```

`./scripts/reproduce_release.sh` remains the frozen paper reproduction path and
must not require Modal credentials.

## Meta OmniASR LLM Unlimited 7B v2

BES-4806 evaluates Meta OmniASR Unlimited 7B v2 as the batch model ID
`modal_meta_omniasr_llm_unlimited_7b_v2`. The checked-in deployment source is
`modal/omniasr_llm_unlimited_7b_v2.py`; do not use the standard 40-second
OmniASR models, the public Replicate endpoint, private deployment code, or a
client-side chunking wrapper.

### OmniASR Deployment Constants

- Benchmark model ID: `modal_meta_omniasr_llm_unlimited_7b_v2`
- Upstream model card: `omniASR_LLM_Unlimited_7B_v2`
- Runtime and catalog revision: `omnilingual-asr==0.2.0`
- Modal app: `voice-code-bench-omniasr`
- Modal GPU: `L40S`
- Modal Volume: `voice-code-bench-omniasr-fairseq2-cache`
- Fairseq2 asset cache: `/root/.cache/fairseq2/assets`
- Language conditioning: `eng_Latn`
- Endpoint secret: `MODAL_OMNIASR_ENDPOINT`

The app mounts the persistent fairseq2 asset cache Volume at
`/root/.cache/fairseq2/assets`, warms the official assets before deployment, and
constructs `ASRInferencePipeline(model_card="omniASR_LLM_Unlimited_7B_v2")` once
per warm container with `@modal.enter`. Each benchmark request sends one
complete 16 kHz mono WAV. There is no client-side chunking, server-side
chunking, prompt, target entity list, custom vocabulary, canonicalization, or
post-ASR correction.

Live OmniASR deployment and release are blocked until the runner has a usable
Modal deployment credential, a deployed Web Function URL, `MODAL_PROXY_TOKEN_ID`,
`MODAL_PROXY_TOKEN_SECRET`, and either a live entity-verifier credential or a
complete verifier cache accepted by `vcb-run`. Mock output is never releasable.

### OmniASR Credentials

Install and configure Modal for the workspace that will own the app:

```bash
python -m pip install modal==1.4.1
modal setup
```

Create a Modal proxy token for the deployment environment:

```bash
modal workspace proxy-tokens create
```

Store only private values in ignored `scripts/.secret` or equivalent environment
variables:

```dotenv
MODAL_OMNIASR_ENDPOINT=https://<your-omniasr-web-function>.modal.run
MODAL_PROXY_TOKEN_ID=wk-...
MODAL_PROXY_TOKEN_SECRET=ws-...
OPENAI_API_KEY=...
```

Never commit `scripts/.secret` or proxy token values.

### Warm The OmniASR Cache

Run cache warm-up before deployment:

```bash
modal run modal/omniasr_llm_unlimited_7b_v2.py::warm_model_cache
```

Verify Modal logs show `omniASR_LLM_Unlimited_7B_v2`,
`omnilingual-asr==0.2.0`, and `/root/.cache/fairseq2/assets`. Rerun the warm-up
if it is interrupted; a successful rerun should reuse the persistent Volume
instead of downloading the full asset set again. If serving reports a missing
asset cache, rerun only this warm-up path before redeploying.

### Deploy OmniASR

Deploy the authenticated Web Function:

```bash
modal deploy modal/omniasr_llm_unlimited_7b_v2.py
```

Copy the Web Function URL printed by Modal, or find it in the dashboard for
`voice-code-bench-omniasr`, and save it as `MODAL_OMNIASR_ENDPOINT`. The
benchmark adapter accepts the URL and normalizes trailing slashes.

The endpoint accepts JSON only:

```json
{
  "audio_base64": "<base64-encoded 16 kHz mono WAV bytes>",
  "audio_format": "wav",
  "language": "eng_Latn"
}
```

The response body is exactly:

```json
{
  "transcript": "<verbatim model output>"
}
```

Requests must include the proxy-token headers:

```text
Modal-Key: <MODAL_PROXY_TOKEN_ID>
Modal-Secret: <MODAL_PROXY_TOKEN_SECRET>
```

A request without those headers should be rejected before inference. Endpoint
errors must not include proxy token values.

### OmniASR Longest-Recording Smoke Test

Create a one-row metadata file for the longest recording,
`audio_id=education_workplace_006` (`data/audio/261.wav`, 122.875 seconds). Do
not use `--limit 1`, because that selects the first metadata row rather than the
longest recording.

```bash
mkdir -p runs/omniasr-longest-smoke
python - <<'PY'
import json
from pathlib import Path
source = Path('data/metadata.jsonl')
out = Path('runs/omniasr-longest-smoke/metadata.longest.jsonl')
for line in source.read_text().splitlines():
    row = json.loads(line)
    if row['audio_id'] == 'education_workplace_006':
        out.write_text(json.dumps(row, separators=(',', ':')) + '\n')
        break
else:
    raise SystemExit('education_workplace_006 not found')
PY
```

Run the smoke transcription:

```bash
vcb-transcribe \
  --metadata runs/omniasr-longest-smoke/metadata.longest.jsonl \
  --stt-model-ids modal_meta_omniasr_llm_unlimited_7b_v2 \
  --output-dir runs/omniasr-longest-smoke
```

Inspect
`runs/omniasr-longest-smoke/modal_meta_omniasr_llm_unlimited_7b_v2/transcripts.json`
and Modal logs. Confirm exactly one non-empty `model_transcript` for
`education_workplace_006`, one complete 122.875-second request, language
`eng_Latn`, batch size 1, and no client-side chunking. Send a second
authenticated smoke request while the container is warm and confirm logs do not
show another pipeline construction for the same warm container.

### OmniASR Full Run And Release Artifact

After the longest-recording smoke test succeeds, run all 300 recordings with
resume enabled from the first invocation:

```bash
vcb-run \
  --stt-model-ids modal_meta_omniasr_llm_unlimited_7b_v2 \
  --resume \
  --output-dir runs/omniasr-full
```

If interrupted, rerun the identical command with `--resume`. Do not regenerate
completed rows. Do not score or release mock transcripts.

Before copying anything to `baselines/predictions/`, validate the final artifact
against `data/metadata.jsonl`: exactly 300 rows, exactly 300 unique `audio_id`
values, identical order, one non-empty `model_transcript` for every row, entity
matches for every target entity, and run metadata containing provider `modal`,
model `omniASR_LLM_Unlimited_7B_v2`, revision `omnilingual-asr==0.2.0`,
`language="eng_Latn"`, the real normalized endpoint, evaluation date, mode
`batch`, L40S, batch size 1, timeout 600 seconds, and 10 max attempts.

Only after that validation passes, copy the generated entity-match artifact to:

```text
baselines/predictions/modal_meta_omniasr_llm_unlimited_7b_v2.json
```

Do not hand-edit transcripts, entity decisions, or run metadata.

Regenerate current aggregate outputs after the real prediction artifact exists:

```bash
vcb-score-entities \
  --dataset-root . \
  --metadata data/metadata.jsonl \
  --entity-matches-dir baselines/predictions \
  --output-dir baselines \
  --scores-dir baselines/scores/entities \
  --output-csv baselines/results.csv

vcb-score-wer \
  --dataset-root . \
  --metadata data/metadata.jsonl \
  --transcripts-dir baselines/predictions \
  --output-dir baselines \
  --scores-dir baselines/scores/wer \
  --output-csv baselines/results.csv

vcb-make-figures --model-set all
```

Update README counts, ranges, and exact OmniASR WER/CTEM/TSR only from the
computed `baselines/results.csv` row. `./scripts/reproduce_release.sh` remains
the frozen paper reproduction path and must not require Modal credentials or add
OmniASR to `PAPER_MODEL_IDS`.

### OmniASR Cleanup

Stop the app when no more requests are expected:

```bash
modal app stop voice-code-bench-omniasr
```

Retain `voice-code-bench-omniasr-fairseq2-cache` while the released baseline may
need reruns. Delete the Volume only when you intentionally want to remove the
cached official assets and are ready to rerun `warm_model_cache`.

## Teardown

After release, stop the app if no more requests are expected:

```bash
modal app stop voice-code-bench-parakeet
```

Retain the `voice-code-bench-parakeet-hf-cache` volume while the released
baseline may need reruns. Delete the volume only when you intentionally want to
remove the cached official weights and are ready to rerun the warm-up step.