cstr commited on
Commit
38de94f
·
verified ·
1 Parent(s): 6269155

Upload README.md with huggingface_hub

Browse files
Files changed (1) hide show
  1. README.md +64 -0
README.md ADDED
@@ -0,0 +1,64 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: afl-3.0
3
+ pipeline_tag: image-to-text
4
+ tags:
5
+ - math
6
+ - ocr
7
+ - handwriting
8
+ - trocr
9
+ - gguf
10
+ - crispembed
11
+ base_model: fhswf/TrOCR_Math_handwritten
12
+ ---
13
+
14
+ # TrOCR Math Handwritten — GGUF
15
+
16
+ Handwritten math equation recognition via [CrispEmbed](https://github.com/CrispStrobe/CrispEmbed)'s ggml inference engine.
17
+
18
+ **Source model**: [fhswf/TrOCR_Math_handwritten](https://huggingface.co/fhswf/TrOCR_Math_handwritten) (AFL-3.0)
19
+
20
+ **Architecture**: ViT-Large encoder (24L, 1024d, 16H) + TrOCR-Large decoder (12L, 1024d, 16H, vocab=50265)
21
+
22
+ ## Models
23
+
24
+ | File | Size | Description |
25
+ |------|------|-------------|
26
+ | `trocr-math-handwritten-fp16.gguf` | 1.2 GB | FP16 weights |
27
+ | `trocr-math-handwritten-q4_k.gguf` | 355 MB | Q4_K quantized |
28
+
29
+ ## Verified Output
30
+
31
+ | Input | Recognized |
32
+ |-------|-----------|
33
+ | x+y=5 (hand-drawn) | `X + Y =` |
34
+ | x²+1 (printed) | `x^{2}` |
35
+ | 3x+7=0 (printed) | `3x+7` |
36
+ | + (cross pattern) | `+` |
37
+
38
+ ## Usage with CrispEmbed
39
+
40
+ ```c
41
+ math_ocr_context* ctx = math_ocr_init("trocr-math-handwritten-q4_k.gguf", 4);
42
+ const char* latex = math_ocr_recognize(ctx, gray_pixels, width, height, &len);
43
+ math_ocr_free(ctx);
44
+ ```
45
+
46
+ ## Conversion
47
+
48
+ Converted from HuggingFace safetensors using CrispEmbed's torch-free converter:
49
+
50
+ ```bash
51
+ python models/convert-trocr-safetensors-to-gguf.py \
52
+ --model-dir /path/to/fhswf/TrOCR_Math_handwritten \
53
+ --output trocr-math-handwritten-fp16.gguf --fp16
54
+ ```
55
+
56
+ Key converter features:
57
+ - No PyTorch dependency (uses `safetensors` + `gguf` + `numpy` only)
58
+ - Generates sinusoidal position embeddings (not stored in checkpoint)
59
+ - Squeezes 3D batch dimensions, reshapes 4D conv weights
60
+ - Preserves AFL-3.0 license in GGUF metadata
61
+
62
+ ## License
63
+
64
+ Academic Free License v3.0 (inherited from source model)