README.md

metadata

license: other
task_categories:
  - question-answering
  - text-generation
language:
  - en
tags:
  - document-qa
  - ocr
  - extractive-qa
  - nanochat
  - sft
pretty_name: DocVQA for Nanochat
size_categories:
  - 10K<n<100K
source_datasets:
  - pixparse/docvqa-single-page-questions

DocVQA for Nanochat

Single-page document QA dataset processed for nanochat fine-tuning.

Description

This dataset is derived from pixparse/docvqa-single-page-questions and has been processed for efficient fine-tuning of small language models with limited context windows.

Modifications from Source

• OCR truncation: Answer-priority truncation ensures the answer is always present in the truncated context. Lines containing the answer are prioritized, then surrounding context is added until the token budget is reached.
• Page numbers: Added "Page X" header at the top of each document from other_metadata.ucsf_document_page_no
• Token budget: Documents truncated to fit within 1750 tokens (for 2048 context window models)
• Short answers: Filtered to answers ≤150 characters
• Format: Conversation format compatible with nanochat's CustomJSON task loader

Statistics

Split	Samples	Total Tokens	Avg Tokens
Train	39,455	15,495,380	393
Validation	5,349	2,218,651	415
Total	44,804	17,714,031	-

Tokenizer

Token counts computed with tiktoken cl100k_base (GPT-4's tokenizer). This is a GPT-4 style BPE tokenizer similar to what nanochat uses.

Schema

Field	Type	Description
question_id	int	Original question ID from DocVQA
question	str	The question to answer
answer	str	The extracted answer (or "Not found in document.")
document_text	str	OCR text with page number prepended
page	int	Page number from OCR results (always 1 for single-page)
other_metadata	dict	Full metadata from source (ucsf_document_id, doc_id, etc.)
num_tokens	int	Exact token count (tiktoken cl100k_base)
match_type	str	How answer was matched: "exact", "fuzzy", or "none"
messages	list	Conversation format for training

Usage

With HuggingFace Datasets

from datasets import load_dataset

ds = load_dataset("morgan/docvqa-nanochat")

# Access a sample
sample = ds["train"][0]
print(f"Question: {{sample['question']}}")
print(f"Answer: {{sample['answer']}}")
print(f"Tokens: {{sample['num_tokens']}}")

For Nanochat Training

The messages field is formatted for nanochat's CustomJSON task:

# Download and convert to JSONL
from datasets import load_dataset
import json

ds = load_dataset("morgan/docvqa-nanochat", split="train")
with open("docvqa_train.jsonl", "w") as f:
    for row in ds:
        f.write(json.dumps(row["messages"]) + "\n")

# Then use with CustomJSON
from tasks.customjson import CustomJSON
train_ds = CustomJSON(filepath="docvqa_train.jsonl")

Document Format

Each document is formatted as:

Document:
Page 4
R. J. REYNOLDS TOBACCO COMPANY
RETAIL PARTNERS MARKETING PLAN CONTRACT
...

Question: When is the contract effective date?

License

Same as source dataset (pixparse/docvqa-single-page-questions).

Citation

If you use this dataset, please cite the original DocVQA paper:

@inproceedings{mathew2021docvqa,
  title={DocVQA: A Dataset for VQA on Document Images},
  author={Mathew, Minesh and Karatzas, Dimosthenis and Jawahar, CV},
  booktitle={WACV},
  year={2021}
}