Alright7398/modded-distill-wavlm-base

---
language:
- en
pretty_name: Modded-DistilWavLM — LibriSpeech + WavLM-Base teacher cache (Lance)
license: cc-by-4.0
tags:
- audio
- speech
- librispeech
- representation-learning
- distillation
- lance
task_categories:
- feature-extraction
size_categories:
- 1M<n<10M
---

# Dataset Summary

Lance tables of **LibriSpeech** utterances at **16 kHz** with **cached last-layer representations** from frozen [`microsoft/wavlm-base`](https://huggingface.co/microsoft/wavlm-base).

This release is a **precomputed teacher cache** plus raw audio bytes, not a general-purpose speech benchmark split.

## Structure

Local / mirrored Hub layout:

| Directory | Content |
|-----------|---------|
| `train/` | LibriSpeech **960 h** train corpora: `train-clean-100`, `train-clean-360`, `train-other-500`. |
| `eval/` | LibriSpeech **dev-clean**. |

Each of `train/` and `eval/` is a **Lance dataset root** (versioned Lance layout on disk), not a single Hugging Face `datasets` Arrow shard tree.

## Schema (per row)

| Column | Type | Description |
|--------|------|-------------|
| `key` | string | Utterance key / basename. |
| `split` | string | Tag written at export (partition **train** vs **eval** is by **top-level folder**, not only this field). |
| `audio` | blob | Raw bytes read from source FLAC (**16 kHz**). |
| `audio_len` | uint32 | Waveform length in samples. |
| `teacher_last` | float16 list | **Flattened** last hidden states from WavLM-Base for valid frames only (length \(T \times H\) for hidden size \(H\)); bf16 forward on GPU, stored as **float16**. |

## Teacher definition (fixed for comparability)

- **Checkpoint**: `microsoft/wavlm-base`.
- **Tensor**: `last_hidden_state`, cropped to **valid** length per utterance (same masking logic as the prepare pipeline), concatenated to a 1D float16 list per row.

## Loading

Requires a **Lance** reader (e.g. Python `lance` / `pylance`):

```python
import lance

ds = lance.dataset("train") # path to train/ after download
print(ds.schema)
print(ds.count_rows())

@inproceedings{panayotov2015librispeech,
title={LibriSpeech: an ASR corpus based on public domain audio books},
author={Panayotov, Vassil and Chen, Guoguo and Povey, Daniel and Khudanpur, Sanjeev},
booktitle={IEEE ICASSP},
year={2015}
}

@inproceedings{chen2022wavlm,
title={WavLM: Large-Scale Self-Supervised Pre-Training for Full Stack Speech Processing},
author={Chen, Sanyuan and others},
booktitle={IEEE JSTSP},
year={2022}
}