Can you recommend a domain-specific LLM for structured extraction from invoices with a small dataset?

Yes — for invoice structured extraction with a small dataset, I’d recommend not starting with a fully domain-trained LLM from scratch. Instead, use a strong general document model or an OCR + extraction model, then fine-tune lightly or use prompting/validation.

Best practical options

1) Donut (NAVER)

• Best when: invoices are images/PDFs and you want end-to-end extraction without a separate OCR step.
• Why it fits small data: it can be fine-tuned on a relatively small labeled set compared to training a model from scratch.
• Strengths: good for document understanding, key-value extraction, table-ish structures.
• Tradeoff: may need careful formatting of labels/output.

2) LayoutLMv3

• Best when: you have OCR text + bounding boxes and want strong structured extraction.
• Why it fits small data: fine-tuning works well on limited labeled data if your OCR is decent.
• Strengths: excellent for invoices, receipts, forms; leverages layout information.
• Tradeoff: requires OCR pipeline and bounding boxes.

3) LiLT / DocFormer

• Best when: your data is mostly text + layout and you want lightweight fine-tuning.
• Strengths: strong document understanding with fewer resources than some larger models.
• Tradeoff: may be less turnkey than Donut.

4) Specialized commercial/document AI models

If you want something ready-made:

• Azure Document Intelligence
• Google Document AI
• Amazon Textract
• ABBYY Vantage These are often the fastest route for invoice extraction with small datasets, especially if you need production reliability.

My recommendation by scenario

If you have:

• Few hundred labeled invoices
• Images/PDFs
• Want a model you can fine-tune end-to-end
  → Donut

If you already have OCR and box coordinates:

• Few hundred to a few thousand labeled examples → LayoutLMv3

If you need the fastest production result:

→ Azure Document Intelligence / Google Document AI / Amazon Textract

Suggested strategy for small data

For invoices, a strong pipeline is often:

1. OCR
2. Field schema definition
   e.g. invoice number, vendor name, date, subtotal, tax, total, line items
3. Fine-tune a document model
4. Post-processing rules
   • date normalization
   • currency parsing
   • total consistency checks
5. Human review for low-confidence cases

If you want an actual LLM-style option

If you specifically want an instruction-tuned LLM for extraction:

• Llama 3.1 / Mistral / Qwen + document OCR text
• Then use structured output prompting or JSON schema constrained decoding This can work well, but for invoices, layout-aware models usually outperform general LLMs unless you have very good prompts and validation.

Bottom line

If I had to pick one for a small labeled invoice dataset:

• Images/PDFs: Donut
• OCR + layout available: LayoutLMv3
• Fastest enterprise solution: Azure Document Intelligence / Google Document AI / Textract

If you want, I can also recommend:

1. a specific open-source model checkpoint,
2. a training pipeline, or
3. the best option based on your dataset size and input format.