---
license: cc0-1.0
language:
  - en
  - hi
  - gu
tags:
  - thinkspark-v2-350m
  - phase1
  - mimi
  - audio-tokens
  - speech
pretty_name: ThinkSpark-v2-350M Phase-1 Free-Audio Corpus
---

# ThinkSpark-v2-350M — Phase-1 free-audio training data

Pre-encoded [Mimi](https://huggingface.co/kyutai/mimi) `cb0` (12.5Hz semantic) tokens +
per-frame `energy`/`f0` prosody, packaged for Phase 1 of ThinkSpark-v2-350M — teaching a
270M-parameter Gemma-3-based full-duplex floor-controller that a stream of Mimi audio
tokens carries language + prosody, before Phase 2 teaches it to referee turn-taking.

Sourced from free/open corpora (LibriSpeech, AI4Bharat Kathbath/Shrutilipi, IndicTTS,
Google FLEURS) via `scripts/P1_00_sequential.py` / `scripts/P1_00_pipeline.py` in the
[kupe-thinkspark-v2-270m](https://github.com/iNavLabsResearch/kupe-thinkspark-v2-270m)
repo — see that repo's `configs/phase1_corpus.yaml` for the exact per-language source
mix, weights, and citations.

## Layout

Each language has its own handful of self-contained Parquet shards:

```
data/en/en-shard-00000.parquet
data/en/en-shard-00001.parquet
...
data/hi/hi-shard-00000.parquet
...
data/gu/gu-shard-00000.parquet
...
```

Every row is one clip's full frame record — audio tokens included directly as columns,
no separate file to join against:

| column | type | meaning |
|---|---|---|
| `scenario_id` | string | stable clip id |
| `behaviour` | string | always `"phase1_free_audio"` here |
| `language` | string | `en` / `hi` / `gu` |
| `domain` | string | source dataset id (e.g. `librispeech`, `kathbath`) |
| `agent_text` | string | always empty for Phase-1 (no agent turn) |
| `user_text` | string | the clip's transcript |
| `num_frames` / `audio_frames` | int | frame count (also `len(cb0)`) |
| `cb0` | list<int64> | Mimi codebook-0 token id, one per 80ms frame |
| `energy` | list<float32> | log-RMS energy per frame |
| `f0` | list<float32> | fundamental frequency (Hz), 0 = unvoiced, per frame |
| `flags` | list<int32> | control-flag id per frame (vocab.CONTROL_FLAG_TO_ID) |
| `agent_state` | list<int32> | agent-state id per frame (vocab.AGENT_STATE_TO_ID) |
| `speaking_mask` | list<int32> | 1 = user speaking that frame |
| `spoken_spans` | list<string> | JSON-encoded spans (empty for Phase-1) |

## Loading

Straight into a training machine's local layout (recommended — no extra code needed,
matches what `scripts/06_train_phase1.py` in the source repo reads directly):
```bash
python scripts/19_fetch_training_data.py --phase1-repo anuj-inavlabs/kupe-thinkspark-270m-phase1-data
```

Or directly via `datasets`/`pandas`/`pyarrow` if you just want to inspect it:
```python
from datasets import load_dataset
ds = load_dataset("anuj-inavlabs/kupe-thinkspark-270m-phase1-data", data_files="data/en/*.parquet", split="train")
```

## License

CC0 — free/open source audio only (see the source repo's `configs/phase1_corpus.yaml`
for each source's own license/citation).