Datasets:

MikhailRudenko
/

domain-aware-sd-synthetic

Tasks:

Languages:

Size:

Tags:

knowledge-distillation

flan

License:

Dataset card Files Files and versions

xet

Community

Dataset Viewer

The dataset could not be loaded because the splits use different data file formats, which is not supported. Read more about the splits configuration. Click for more details.

Couldn't infer the same data file format for all splits. Got {NamedSplit('train'): (None, {}), NamedSplit('validation'): ('json', {}), NamedSplit('test'): (None, {})}

Error code:   FileFormatMismatchBetweenSplitsError

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

Domain-Aware Speculative Decoding — Synthetic Dataset

Synthetic distillation data for training domain-specific draft models in speculative decoding.

Overview

Generated by running TurboSparse-Mistral-Instruct (7B) in teacher-forcing mode on Flan tasks and capturing top-10 token distributions at each position.

Structure

train/v3/    — 66 NPZ files (2.2 GB), one per Flan cluster
validation/v3/ — 66 JSONL files (93 MB)
test/v3/     — 66 NPZ files (598 MB)

Record Schema (JSONL)

{
  "cluster": "aeslc_10templates",
  "prompt": "<flan input>",
  "reference": "<flan target>",
  "trunk": [1234, 567, ...],
  "top10_ids": [[1234, 99, 12, ...], ...],
  "top10_probs": [[0.45, 0.20, 0.05, ...], ...]
}

NPZ Format

Same fields stored as numpy arrays. See DATASET_FORMAT.md.

Target Model

TurboSparse-Mistral-Instruct (7B, BambooForCausalLM)
Vocab: 32064 tokens (Mistral base + 64 special)

Usage

from huggingface_hub import snapshot_download
snapshot_download("MikhailRudenko/domain-aware-sd-synthetic", repo_type="dataset", local_dir="data/synthetic")

Downloads last month: 336

Total file size:

2.99 GB

Models trained or fine-tuned on MikhailRudenko/domain-aware-sd-synthetic