Qwen2.5-VL-3B-Instruct
A 3-billion-parameter vision-language model from Alibaba's Qwen2.5 series, supporting multimodal image-text understanding and instruction following.
Timeline (last 90 days)
The learning curve shows that vertical writing layouts cannot be solved by adding data: 8.7% with zero training samples, 84.7% with 800 sheets, and 89.3% for both 2400 and 3000 sheets.
Without validation, 12.8% of the fine-tuned model's invoices have validation-coverage field errors.
Horizontal layouts perform excellently: 97.3% exact on the test set and 99.0% on unseen horizontal layouts; vertical writing is the weak point: 40.0% on the test set, mainly due to Latin text printed sideways (registration number 62.7%, date 82.7%), while the table portion is normal; in unseen fully vertical layouts the table is rotated, at only 4%.
Developer raihan-js used Qwen2.5-VL-3B-Instruct with a QLoRA adapter, fine-tuning for 1 epoch on 3000 synthetic invoices, taking 4 hours 57 minutes on a single RTX 3060.
On a test set of 600 synthetic invoices, the QLoRA fine-tuned model achieved a full-invoice exact rate of 83.0%, an auto-approval rate of 86.8%, and 0 out of 461 validation-coverage field errors in approved invoices (upper bound 0.8%).
Zero-shot Qwen2.5-VL-3B-Instruct achieved a full-invoice exact rate of 5.3% and an auto-approval rate of 8.1% on the test set; the OCR + rules baseline achieved a full-invoice exact rate of 37.8% and an auto-approval rate of 55.0%.
Qwen2.5-3B was excluded because its license is marked as other, and Qwen3-8B was excluded because it exceeds memory; DeepSeek-R1-Distill-Qwen-1.5B and OLMo 2 1B, although they meet the license requirements, cannot be loaded on this vLLM image.
The researchers validated the parity conclusion on the vision-language model Qwen2.5-VL-3B-Instruct for a single compression ratio using per-scheme learning rate tuning repeated with three random seeds.
All Facts (8)
The learning curve shows that vertical writing layouts cannot be solved by adding data: 8.7% with zero training samples, 84.7% with 800 sheets, and 89.3% for both 2400 and 3000 sheets.
50%UnverifiedWithout validation, 12.8% of the fine-tuned model's invoices have validation-coverage field errors.
50%UnverifiedHorizontal layouts perform excellently: 97.3% exact on the test set and 99.0% on unseen horizontal layouts; vertical writing is the weak point: 40.0% on the test set, mainly due to Latin text printed sideways (registration number 62.7%, date 82.7%), while the table portion is normal; in unseen fully vertical layouts the table is rotated, at only 4%.
50%UnverifiedDeveloper raihan-js used Qwen2.5-VL-3B-Instruct with a QLoRA adapter, fine-tuning for 1 epoch on 3000 synthetic invoices, taking 4 hours 57 minutes on a single RTX 3060.
50%UnverifiedOn a test set of 600 synthetic invoices, the QLoRA fine-tuned model achieved a full-invoice exact rate of 83.0%, an auto-approval rate of 86.8%, and 0 out of 461 validation-coverage field errors in approved invoices (upper bound 0.8%).
50%UnverifiedZero-shot Qwen2.5-VL-3B-Instruct achieved a full-invoice exact rate of 5.3% and an auto-approval rate of 8.1% on the test set; the OCR + rules baseline achieved a full-invoice exact rate of 37.8% and an auto-approval rate of 55.0%.
50%UnverifiedQwen2.5-3B was excluded because its license is marked as other, and Qwen3-8B was excluded because it exceeds memory; DeepSeek-R1-Distill-Qwen-1.5B and OLMo 2 1B, although they meet the license requirements, cannot be loaded on this vLLM image.
50%UnverifiedThe researchers validated the parity conclusion on the vision-language model Qwen2.5-VL-3B-Instruct for a single compression ratio using per-scheme learning rate tuning repeated with three random seeds.
50%