---
language:
  - en
  - fr
license: apache-2.0
task_categories:
  - question-answering
  - translation
  - text-retrieval
tags:
  - long-context
  - prefill
  - tokenized
  - qwen3
  - kv-cache
  - benchmark
pretty_name: Prefill Dataset - Long-Context Tokenized Corpus for Qwen3-8B
size_categories:
  - 1K<n<10K
---

# Prefill Dataset

Long-context tokenized corpus for benchmarking LLM prefill computation with Qwen3-8B. Contains ~10M tokens of copyright-free English text pre-tokenized with character offset mappings for fast position lookup.

## Dataset Structure

### Files

| File | Description | Rows |
|------|-------------|------|
| `data/documents.parquet` | English documents with token IDs and char offsets | ~100-500 |
| `data/tasks.parquet` | QA, translation, and retrieval tasks | ~1K-5K |
| `data/translations.parquet` | French translations of OPUS-Books English documents | ~100-500 |
| `data/aligned_chunks.parquet` | EN/FR aligned chunk pairs packed to ~1k source tokens | ~1K-5K |

### `documents.parquet` Schema

| Column | Type | Description |
|--------|------|-------------|
| `doc_id` | string | Unique document ID |
| `source` | string | `"narrativeqa"` / `"opus_books"` / `"pg19"` |
| `title` | string | Book title |
| `language` | string | Always `"en"` |
| `text` | large_string | Full document text |
| `token_ids` | list\<int32\> | Qwen3-8B token IDs |
| `char_offsets` | list\<int32\> | Character start position per token |
| `token_count` | int32 | Number of tokens |

### `tasks.parquet` Schema

| Column | Type | Description |
|--------|------|-------------|
| `task_id` | string | Unique task ID |
| `doc_id` | string | References `documents.doc_id` |
| `task_type` | string | `"qa"` / `"translation"` / `"retrieval"` |
| `question` | string | Task prompt |
| `answer` | string | Expected answer (JSON list for multi-answer) |
| `metadata` | string | JSON with extra fields |

### `translations.parquet` Schema

| Column | Type | Description |
|--------|------|-------------|
| `doc_id` | string | References English document |
| `target_language` | string | Always `"fr"` |
| `target_text` | large_string | Full translation text |
| `target_token_ids` | list\<int32\> | Tokenized translation |
| `target_char_offsets` | list\<int32\> | Char offsets for translation tokens |

### `aligned_chunks.parquet` Schema

| Column | Type | Description |
|--------|------|-------------|
| `chunk_id` | string | Unique chunk ID (`doc_id` + chunk index) |
| `doc_id` | string | References OPUS English document |
| `chunk_idx` | int32 | Chunk index within document |
| `segment_start_idx` | int32 | Start aligned segment index (inclusive) |
| `segment_end_idx` | int32 | End aligned segment index (exclusive) |
| `src_lang` | string | Always `"en"` |
| `tgt_lang` | string | Always `"fr"` |
| `src_text` | large_string | English chunk text |
| `tgt_text` | large_string | French chunk text |
| `src_char_start` / `src_char_end` | int32 | Character span in source document |
| `tgt_char_start` / `tgt_char_end` | int32 | Character span in translation document |
| `src_tok_start` / `src_tok_end` | int32 | Token span in source token IDs |
| `tgt_tok_start` / `tgt_tok_end` | int32 | Token span in target token IDs |
| `src_token_count` | int32 | Source tokens in chunk (target ~1000) |
| `tgt_token_count` | int32 | Target tokens in chunk |

## Sources

| Source | Purpose | Target Tokens |
|--------|---------|---------------|
| [NarrativeQA](https://huggingface.co/datasets/deepmind/narrativeqa) | Gutenberg books with human Q&A pairs | ~5M |
| [OPUS-Books](https://huggingface.co/datasets/Helsinki-NLP/opus_books) | Parallel EN-FR book translations | ~3M |
| [pg19](https://huggingface.co/datasets/deepmind/pg19) | Supplementary long Gutenberg books | ~2M+ |

## Tokenizer

- **Model**: `Qwen/Qwen3-8B` (vocab size: 151,936)
- **Offset mapping**: `char_offsets[i]` is the character position where token `i` starts. BPE tokens with leading spaces point to the space character — this is correct: `text[char_offsets[i]:char_offsets[i+1]]` recovers exact token text.

## Usage

```python
import pyarrow.parquet as pq

# Load
docs = pq.read_table("data/documents.parquet").to_pandas()
tasks = pq.read_table("data/tasks.parquet").to_pandas()

# Get a document and its tasks
doc = docs.iloc[0]
doc_tasks = tasks[tasks.doc_id == doc.doc_id]

print(f"Title: {doc.title}")
print(f"Tokens: {doc.token_count:,}")
print(f"Tasks: {len(doc_tasks)}")

# Verify token-to-text mapping
offsets = doc.char_offsets
text = doc.text
for i in range(5):
    end = offsets[i + 1] if i + 1 < len(offsets) else len(text)
    print(f"  Token {i}: '{text[offsets[i]:end]}'")
```

See `generate_examples.py` for a full usage example.

## Regeneration

```bash
uv run build_dataset.py
```

Requires Python 3.11+. Dependencies are declared inline (PEP 723) — `uv run` handles them automatically.

## License

The dataset is released under Apache 2.0. Source texts are public domain (Project Gutenberg).