Typhoon OCR 7B — Thai Handwriting LoRA LoRA adapter on typhoon ai/typhoon ocr 7b , fine tuned to read Thai handwriting from an image and return the transcription as JSON ( {"natural text": "..."} ). Overview Base model: typhoon ai/typhoon ocr 7b (Qwen2.5 VL 7B, Apache 2.0) Task: Thai handwriting image → text Type: PEFT/LoRA adapter — load on top of the base model (base weights are not redistributed here) Intended inference: images at 1024px ( max pixels=1048576 ), greedy decoding ( do sample=False ), max new tokens=512 . Other regimes are outside the evaluation scope. Performance This is the evaluation attached to the branch promoted to main : phase2 downsample money keep25 300steps , evaluated on the 559 sample CPE OPH full test with 1024px images, greedy decoding, and max new tokens=512 . Metric Promotion baseline Promoted main Delta Direction : : : valid json rate 1.0000 1.0000 +0.0000 higher is better raw cer 0.1560 0.1336 0.0224 lower is better normalized cer 0.1560 0.1336 0.0224 lower is better thai char accuracy 0.8440 0.8664 +0.0224 higher is better exact match rate 0.4204 0.4526 +0.0322 higher is better Baseline: phase2 baseline greedy res1024 . Source: published metrics a…
We use cookies for essential functionality and analytics. You can accept or reject analytics cookies.Cookie policy