tiny-random/minicpm-ocr-jev

This is a randomly initialized tiny model based on openbmb/MiniCPM-OCR-Jev for testing and debugging.

Feature	Source	This repo
Checkpoint tensor elements	2.21B BF16	6.64M BF16
Safetensors weight files	4.43 GB (model.safetensors)	13.31 MB (model.safetensors)
Architecture / input	Qwen3.5 vision-language correctness scorer	Same
Vocabulary rows / tied embeddings	248,320 / false	Same
Backbone layers / layer cadence	24 / six 3-linear + 1-full cycles	8 / two complete cycles
Hidden width / MLP width	2,048 / 6,144	16 / 32
Q heads / KV heads / head dimension	8 / 2 / 256	Same
Linear-attention key/value heads and dimensions	16/16, 128/128	Same
Vision depth / width / output width	24 / 1,024 / 2,048	4 / 128 / 16
Vision patch / temporal patch / spatial merge	16 / 2 / 2	Same
Task head	binary score [2, 2,048]; no lm_head	binary score [2, 16]; no lm_head
MTP	config declares one layer; no MTP tensors	Same
Checkpoint precision	BF16	Same

Load this checkpoint with AutoModelForSequenceClassification; it keeps the source binary score head and does not contain lm_head. Two complete linear/linear/linear/full cycles are retained. The Qwen3.5 linear-attention path requires CUDA and Triton. The exact public snippet was verified on an NVIDIA A800-SXM4-80GB.

File path	Size
model.safetensors	13.3MB

Example usage

Transformers (✅ verified)

The example sends two images: the original block and a rendered image of the candidate text. Load with AutoModelForSequenceClassification; the checkpoint has a binary score head instead of lm_head. Install transformers>=5.10.4 and Pillow; the Qwen3.5 linear-attention path requires CUDA and Triton.

def run_transformers(model_id: str):
    import torch
    from PIL import Image, ImageDraw
    from transformers import AutoModelForSequenceClassification, AutoProcessor

    model_id = str(model_id)
    processor = AutoProcessor.from_pretrained(model_id)
    model = AutoModelForSequenceClassification.from_pretrained(
        model_id,
        dtype=torch.bfloat16,
        device_map="cuda",
        attn_implementation="sdpa",
    )
    candidate = "Hello"
    original = Image.new("RGB", (448, 448), "white")
    ImageDraw.Draw(original).text((20, 20), candidate, fill="black")
    rendered = Image.new("RGB", (448, 448), "white")
    ImageDraw.Draw(rendered).text((20, 20), candidate, fill="black")
    instruction = (
        "第一张图是原图，第二张图是上述 OCR 文本渲染得到的图片。\n"
        "请判断该 OCR 文本及其对应渲染图是否与原图内容一致。\n"
        "这是一个二分类任务：一致标为 1，不一致标为 0。"
    )
    messages = [
        {
            "role": "user",
            "content": [
                {"type": "image", "image": original},
                {"type": "text", "text": f"\nOCR文本：\n{candidate}\n"},
                {"type": "image", "image": rendered},
                {"type": "text", "text": "\n" + instruction},
            ],
        }
    ]
    inputs = processor.apply_chat_template(
        messages,
        tokenize=True,
        add_generation_prompt=True,
        return_dict=True,
        return_tensors="pt",
    ).to(model.device)
    with torch.no_grad():
        logits = model(**inputs).logits
        assert torch.isfinite(logits).all()
        score = torch.softmax(logits.float(), dim=-1)[:, 1]
    print(float(score[0]))


if __name__ == "__main__":
    run_transformers(model_id="tiny-random/minicpm-ocr-jev")

Creating this repo

• Code: .meta/create_model.py
• Model structure: .meta/model.log
• Environment: .meta/env.json

Change log

• 2026-11-01: Initial tiny MiniCPM-OCR Jev checkpoint.

This repo is generated by DeepSeek-V4.1-Flash.