4 ms·
I’ve tried that too, trying to detect the scan layout to get better OCR, but it didn’t really beat a fine-tuned Qwen 2.5 VLM 7B. I’d say fine-tuning is the way
by netdur 1y ago
I’ve tried that too, trying to detect the scan layout to get better OCR, but it didn’t really beat a fine-tuned Qwen 2.5 VLM 7B. I’d say fine-tuning is the way to go
- rexreed 1y agowhat fine tuning approach did you use?
- netdur 1y agojust unsloth on colab using A100 and dataset on google drive.
- richardlblair 1y agoWhat's the cost of the fine-tuned model? If you were attempting to optimize for cost, would it be worth it to detect scan layouts to get better OCR? Honestly, I'm such a noob in this space. I had 1 project I needed to do, didn't want to do it by hand which would have taken 2 days so I spent 5 trying to get a script to do it for me.
- netdur 1y agothe model runs on H200 in ~20s, costing about $2.4/hr. on L4 it’s cheaper at ~$0.3/hr but takes ~85s to finish. overall, H200 ends up cheaper at volume. my scan has a separate issue though: each page has two columns, so text from the right side sometimes overflows into the left. OCR can’t really tell where sentences start and end unless the layout is split by column.