Possible numeric precision drift from official demo
I have an image in which table can not be parsed correctly with llama.cpp(R-SWA branch). But it can be parsed correctly in the official demo. I tried variable quantized models. Even both bf16 combination can not work.
The table 2 ocr result is like this:
| Methods | " | "" | ||
| TEDs-S | TEDS | > | ||
| TableFormer[2022] | 96.80 | - | 96.70 | |
| GridFormer [2023] | 98.63 | - | - | |
| VAST[2023] | 98.633 | 98.21 | - | |
| TFLO \( P_{BASE} \) | 99.43 | 99.22 | 99.42 | 99.34 |
| TFLO \( P_{FULL} \) | 99.56 | 99.45 | 99.42 | \( \underline{99.40} \) |
Thanks for the report - reproduced and fixed.
Not a precision issue: the DRY flags in my run example caused it. They punish repetition, and the model writes tables as HTML markup - one <td> tag per cell, so lots of legitimate repetition. I removed the flags from the run examples (README + PRs); with the current command your image parses correctly and matches the HF reference.
Your image is now a regression case in https://github.com/sfallah/llama-mtmd-eval, so this can't come back silently.
If you want to verify with your image:
build/bin/llama-mtmd-cli -hf sabafallah/Unlimited-OCR-GGUF:bf16 \
--image your-image.png -p "document parsing." \
--chat-template deepseek-ocr \
--temp 0 --flash-attn off --no-warmup \
-n 8192 -c 16384 \
--dry-multiplier 0.8 --dry-base 1.75 --dry-allowed-length 35 \
--dry-penalty-last-n 128 --dry-sequence-breaker none
These DRY settings mirror the reference's no_repeat_ngram guard (35 / window 128) and are safe for tables - the old aggressive ones (allowed-length 2) were the culprit. The README run example is updated accordingly.