Instructions to use enclavelabs/olmocr-2-iter7a-vqa with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use enclavelabs/olmocr-2-iter7a-vqa with PEFT:

from peft import PeftModel
from transformers import AutoModelForCausalLM

base_model = AutoModelForCausalLM.from_pretrained("allenai/olmOCR-2-7B-1025")
model = PeftModel.from_pretrained(base_model, "enclavelabs/olmocr-2-iter7a-vqa")

Transformers

How to use enclavelabs/olmocr-2-iter7a-vqa with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("image-text-to-text", model="enclavelabs/olmocr-2-iter7a-vqa")

# Load model directly
from transformers import AutoModel
model = AutoModel.from_pretrained("enclavelabs/olmocr-2-iter7a-vqa", device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use enclavelabs/olmocr-2-iter7a-vqa with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "enclavelabs/olmocr-2-iter7a-vqa"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "enclavelabs/olmocr-2-iter7a-vqa",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/enclavelabs/olmocr-2-iter7a-vqa

SGLang

How to use enclavelabs/olmocr-2-iter7a-vqa with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "enclavelabs/olmocr-2-iter7a-vqa" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "enclavelabs/olmocr-2-iter7a-vqa",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "enclavelabs/olmocr-2-iter7a-vqa" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "enclavelabs/olmocr-2-iter7a-vqa",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use enclavelabs/olmocr-2-iter7a-vqa with Docker Model Runner:
```
docker model run hf.co/enclavelabs/olmocr-2-iter7a-vqa
```

EnclaveScribe iter-7a — Scoped English document VQA specialist

⚠️ NOT a general OCR model. This adapter is scoped to English document VQA and short-answer document tasks. For general OCR — especially anything non-Latin (Devanagari, Chinese, Arabic, …) — use a different model.

Trained by Enclave Labs as part of a public, iteration-by-iteration effort to build a sovereign (fully self-hosted) alternative to Interfaze / Gemini / GPT-4V for document intelligence.

TL;DR

What it is: a LoRA adapter (r=32, α=64) on top of allenai/olmOCR-2-7B-1025.
What it does well: English document VQA. Answers questions asked about a document image. 5.4× the OCRBench V2 F1 of base olmOCR-2-7B-1025.
What it does badly: transcribing Devanagari, or any script it wasn't substantially trained on. Regresses base's already-low Devanagari CER by ~19× — do not use for that.

Benchmark results (iter-8 corrected measurement, 2026-09-14)

Re-measured after fixing an eval bug that hid iter-7a's real OCRBench V2 performance (fixed-prompt "document parsing." was dropping OCRBench V2's per-task questions).

Benchmark	Metric	iter-7a	base olmOCR-2-7B-1025	delta
OCRBench V2 (300 samples, per-sample prompts)	F1 ↑	0.499	0.093	5.4×
OmniDocBench (250 pages, `research_report` @2048 tokens)	F1 ↑	0.293	0.291	+0.7%
himalaya_500 (500 Devanagari word crops)	CER ↓	4.463 (446%) ❌	14.32 (1432%) ❌	—

For OCRBench V2, iter-7a lands at ~71% of Interfaze's 0.707 target and ~78% of Gemini-3.5-Flash's 0.639. It is currently Enclave's strongest result on this benchmark by a wide margin.

For Devanagari, iter-7a is worse than iter-3 (enclavelabs/enclave-scribe-devanagari, CER 0.175) and iter-4 (CER 0.231) by an order of magnitude — do not use.

Recommended use

Good fits:

Answering questions asked about scanned documents ("What is the total amount?", "What date is on this receipt?", "Who is the sender?").
English document short-answer extraction where the model has a specific question to answer.
Downstream fine-tuning on further English document VQA data.

Bad fits — use a different model:

Non-Latin scripts (Devanagari, Chinese, Arabic, Japanese, Korean, …). For Devanagari, use enclavelabs/enclave-scribe-devanagari (iter-3), CER 0.175 on himalaya_500.
Long-form page transcription ("extract this entire page as markdown"). Iter-7a's OmniDocBench F1 (0.293) is only marginally above base olmOCR-7B (0.291); base is essentially as good and doesn't come with the Devanagari regression.
Structural output (bounding boxes, confidence scores, layout). Not trained for it.

Prompting

Feed each sample its actual per-task question in the prompt slot, e.g.:

from transformers import AutoProcessor, AutoModelForVision2Seq
from peft import PeftModel

BASE = "allenai/olmOCR-2-7B-1025"
ADAPTER = "Enclave-Labs-Inc/olmocr-2-iter7a-vqa"

processor = AutoProcessor.from_pretrained(BASE)
model = AutoModelForVision2Seq.from_pretrained(BASE, torch_dtype="bfloat16", device_map="auto")
model = PeftModel.from_pretrained(model, ADAPTER)

messages = [{
    "role": "user",
    "content": [
        {"type": "image", "image": "receipt.jpg"},
        {"type": "text",  "text": "What is the total amount?"},
    ],
}]

Passing the fixed prompt "document parsing." (as some pipelines do by default) will produce weaker results — iter-7a's advantage is on tasks with real per-sample questions.

Training

Base model: allenai/olmOCR-2-7B-1025 (Qwen2.5-VL-7B fine-tuned by AI2 for OCR).
Method: fresh LoRA (r=32, α=64, dropout=0.05), 2 epochs, LR 1.0e-4, bf16, gradient checkpointing, liger kernel, max_pixels=200704 (448²), max_length=4096.
Data (~10k filtered samples): DocVQA (HuggingFaceM4/DocumentVQA, short-answer mode), XFUND (7 langs), IDL-WDS (historical documents), iter-4 Devanagari replay (~500). URL rot dropped the originally-planned TextOCR + HierText sources — see the report for details.
Hardware: g5.4xlarge on-demand (1× A10G 24GB), us-east-1.
Cost: ~$48 as-shipped.

Full details, deviations from plan, and postmortem in reports/iter7a/README.md.

Framework versions

PEFT 0.20.0
transformers 4.55.4
torch 2.7.0+cu128

Iteration lineage

iter-3 — Devanagari word-level specialist (enclavelabs/enclave-scribe-devanagari). 93× CER improvement over base on himalaya_500.
iter-4 — Devanagari page-level (unpublished; superseded by iter-3 for most Devanagari uses because iter-4 regressed slightly).
iter-7a — this model. Scoped English VQA specialist. Do not use for Devanagari.
iter-9 (in progress) — extend iter-7a's recipe with expanded VQA / OCR-VQA corpus, targeting Interfaze / Gemini on OCRBench V2. If successful will be published as Enclave-Labs-Inc/enclave-scribe-iter9.

License

MIT. Same as the base allenai/olmOCR-2-7B-1025.

Citation

If you use this model, please cite the base model and the EnclaveScribe repository:

@misc{enclave-scribe-iter7a-vqa,
  title={EnclaveScribe iter-7a: Scoped English document VQA LoRA on olmOCR-2-7B-1025},
  author={Enclave Labs},
  year={2026},
  url={https://huggingface.co/Enclave-Labs-Inc/olmocr-2-iter7a-vqa}
}

Downloads last month: 17

Model tree for enclavelabs/olmocr-2-iter7a-vqa

Base model

Qwen/Qwen2.5-VL-7B-Instruct

Finetuned

allenai/olmOCR-2-7B-1025

Adapter

(9)

this model