File size: 4,129 Bytes
332fd89 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 | ---
library_name: transformers
pipeline_tag: image-text-to-text
base_model:
- openbmb/MiniCPM-OCR-Jev
---
This is a randomly initialized tiny model based on [openbmb/MiniCPM-OCR-Jev](https://huggingface.co/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.
```python
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](.meta/create_model.py)
- Model structure: [.meta/model.log](.meta/model.log)
- Environment: [.meta/env.json](.meta/env.json)
### Change log
- 2026-11-01: Initial tiny MiniCPM-OCR Jev checkpoint.
*This repo is generated by DeepSeek-V4.1-Flash.* |