Hugging Face's logo

context212
/
alhazen-ocr

Image-Text-to-Text
PEFT
Safetensors
arabic
ocr
qwen3-vl
lora
unsloth
document-ai
conversational
Model card Files
xet
 Community
isaac0x53 commited on
Commit
110b461
·
verified ·
1 Parent(s): b951a9d

Update model card: base vs Alhazen-OCR (khattmix eval)

Browse files
Files changed (1) hide show
  1. README.md +8 -6
README.md CHANGED
@@ -19,9 +19,9 @@ QLoRA adapter on [`unsloth/Qwen3-VL-2B-Instruct`](https://huggingface.co/unsloth
19
  Named after Ibn al-Haytham (Alhazen), the 11th-century scholar who founded the science of optics.
20
 
21
  - **Base:** Qwen3-VL-2B-Instruct (Apache-2.0)
22
- - **Training data:** [`context212/context212-alhazen-ocr`](https://huggingface.co/datasets/context212/context212-alhazen-ocr)
23
  - **Code:** [`github.com/context212/atlas-ocr`](https://github.com/context212/atlas-ocr)
24
- - **Experiment sibling:** [`context212/alhazen-ocr-hwmix`](https://huggingface.co/context212/alhazen-ocr-hwmix) (same weights; promoted here after KHATT improved)
25
 
26
  ## Results
27
 
@@ -30,17 +30,19 @@ CER / WER (lower is better), greedy decoding, **50 samples** each on the held-ou
30
  | Model | Eval CER ↓ | Eval WER ↓ | KHATT CER ↓ | KHATT WER ↓ |
31
  |---|---:|---:|---:|---:|
32
  | Qwen3-VL-2B-Instruct (base) | 0.792 | 0.947 | 1.893 | 1.801 |
33
- | **Alhazen-OCR** | **0.281** | **0.448** | **1.121** | **1.215** |
34
 
35
- Held-out character error drops from **0.79 → 0.28**. On external KHATT handwriting, CER/WER also improve versus the untuned base (**1.89 → 1.12** CER) — the previous synthetic-heavy mix had *regressed* on KHATT; this release uses a handwriting-heavier mix (~45% KHATT paragraphs, ~35% synthetic printed, ~20% invoices, plus a small historical set; KHATT-bench transcript overlap removed from train).
 
 
36
 
37
  ## Training
38
 
39
- - **Objective:** QLoRA SFT, one epoch (second epoch previously diverged)
40
  - **LR:** 2e-5 cosine, warmup 5%
41
  - **Effective batch:** 16 (8 × grad accum 2) on A100 80GB
42
  - **LoRA:** r=16, RSLoRA, vision + language layers
43
- - **Mix:** handwriting-heavy rebuild of `context212/context212-alhazen-ocr` (~15k rows after mix)
44
 
45
  ## Usage
46
 
 
19
  Named after Ibn al-Haytham (Alhazen), the 11th-century scholar who founded the science of optics.
20
 
21
  - **Base:** Qwen3-VL-2B-Instruct (Apache-2.0)
22
+ - **Training data:** [`context212/context212-alhazen-ocr`](https://huggingface.co/datasets/context212/context212-alhazen-ocr) (khattmix rebuild)
23
  - **Code:** [`github.com/context212/atlas-ocr`](https://github.com/context212/atlas-ocr)
24
+ - **Experiment sibling:** [`context212/alhazen-ocr-khattmix`](https://huggingface.co/context212/alhazen-ocr-khattmix)
25
 
26
  ## Results
27
 
 
30
  | Model | Eval CER ↓ | Eval WER ↓ | KHATT CER ↓ | KHATT WER ↓ |
31
  |---|---:|---:|---:|---:|
32
  | Qwen3-VL-2B-Instruct (base) | 0.792 | 0.947 | 1.893 | 1.801 |
33
+ | **Alhazen-OCR** | **0.328** | **0.504** | **0.841** | **1.193** |
34
 
35
+ Held-out character error: **0.79 → 0.33**. External KHATT handwriting: **1.89 → 0.84** CER.
36
+
37
+ This release (khattmix) adds line-level KHATT handwriting (~30% of the mix) on top of synthetic printed, invoices, and paragraph handwriting; KHATT-bench transcript overlap is removed from train.
38
 
39
  ## Training
40
 
41
+ - **Objective:** QLoRA SFT, one epoch
42
  - **LR:** 2e-5 cosine, warmup 5%
43
  - **Effective batch:** 16 (8 × grad accum 2) on A100 80GB
44
  - **LoRA:** r=16, RSLoRA, vision + language layers
45
+ - **Mix:** ~30% loay synthetic / 15% invoices / 30% KHATT lines / 23% KHATT paragraphs / 2% historical (~18k train rows)
46
 
47
  ## Usage
48