vikp

It's an interesting illustration of the state of the AI market that immediately after arguing that AI cannot do anything complex...we have an ad arguing that AI can actually do those things. Even the people telling you that AI is a hype cycle buy into the hype cycle.

First, we have this section:

But the noise continues on volume 11/10. So this is my desperate plea... Can we PUHLEASE stop this AI Confidence Theater, people? It’s doing more harm than good.

Then immediately after:

This post is sponsored by Firecrawl. Firecrawl is the web data API to search, scrape, and interact with the web at scale. Turn the live web into clean, structured data your agents can actually use.

(P.S. I personally use Firecrawl in my Lovable apps all.the.time - most recently to scrape this very blog so my AI double could use my latest posts as context.)

Hey, I'm the founder of Datalab (we released Chandra OCR). I see someone requested it below - happy to help you all get setup. I'm [email protected]

Yes, we can sign a BAA!

Hi, I'm a founder of Datalab. I'm not trying to take away from the launch (congrats), just wanted to respond to the specific feedback.

I'm glad you found a solution that worked for you, but this is pretty surprising to hear - our new model, chandra, saturates handwriting-heavy benchmarks like this one - https://www.datalab.to/blog/saturating-the-olmocr-benchmark ,and our production models are more performant than OSS.

Did you test some time ago? We've made a bunch of updates in the last couple of months. Happy to issue some credits if you ever want to try again - [email protected].

I assume you're using a PDF, and not the image you shared? You need to set force ocr or format lines to get inline math with a PDF (for images, we just OCR everything anyways, so you don't need any settings).

We're working on improving the playground generally now - expect a big update tomorrow, which among other things will default to format lines.

Thanks for the kind words! The team was just me until pretty recently, but we're growing quickly and will be addressing a lot of issues quickly in the next few weeks.

Hi, author of marker here - I tried your image, and I don't see the issues you're describing with the newest version of marker (1.7.5).

I ran both with no setting specified, and with force_ocr, and I didn't see the issues either time.

Thanks for sharing! I'm training some models now that will hopefully improve this and more :)

Benchmarking is hard for markdown because of the slight formatting variations between different providers. With HTML, you can use something like TEDS (although there are issues with this, too), but with markdown, you don't have a great notion of structure, so you're left with edit distance.

I think blockwise edit distance is better than full page (find the ground truth blocks, then infer each block separately and compare), but many providers only do well on full pages, which doesn't make it fair.

There are a few different benchmark types in the marker repo:

  - Heuristic (edit distance by block with an ordering score)
  - LLM judging against a rubric
  - LLM win rate (compare two samples from different providers)

None of these are perfect, but LLM against a rubric has matched visual inspection the best so far.

I'll continue to iterate on the benchmarks. It may be possible to do a TEDS-like metric for markdown. Training a model on the output and then benchmarking could also be interesting, but it gets away from measuring pure extraction quality (the model benchmarking better is only somewhat correlated with better parse quality). I haven't seen any great benchmarking of markdown quality, even at research labs - it's an open problem.

I ran a partial benchmark against marker - https://github.com/VikParuchuri/marker .

Across 375 samples with LLM as a judge, mistral scores 4.32, and marker 4.41 . Marker can inference between 20 and 120 pages per second on an H100.

You can see the samples here - https://huggingface.co/datasets/datalab-to/marker_comparison... .

The code for the benchmark is here - https://github.com/VikParuchuri/marker/tree/master/benchmark... . Will run a full benchmark soon.

Mistral OCR is an impressive model, but OCR is a hard problem, and there is a significant risk of hallucinations/missing text with LLMs.

I'm a fan of the team of Allen AI and their work. Unfortunately, the benchmarking of olmocr against marker (https://github.com/VikParuchuri/marker) is quite flawed.

Throughput - they benchmarked marker API cost vs local inference cost for olmocr. In our testing, marker locally gets 20 - 120 pages per second on an H100 (without custom kernels, etc). Olmocr in our testing gets between .4 (unoptimized) and 4 (sglang) pages per second on the same machine.

Accuracy - their quality benchmarks are based on win rate with only 75 samples - which are different between each tool pair. The samples were filtered down from a set of ~2000 based on opaque criteria. They then asked researchers at Allen AI to judge which output was better. When we benchmarked with our existing set and LLM as a judge, we got a 56% win rate for marker across 1,107 documents. We had to filter out non-English docs, since olmocr is English-only (marker is not).

Hallucinations/other problems - we noticed a lot of missing text and hallucinations with olmocr in our benchmark set. You can see sample output and llm ratings here - https://huggingface.co/datasets/datalab-to/marker_benchmark_... .

You can see all benchmark code at https://github.com/VikParuchuri/marker/tree/master/benchmark... .

Happy to chat more with anyone at Allen AI who wants to discuss this. I think olmocr is a great contribution - happy to help you benchmark marker more fairly.

Docling is a great project, happy to see more people building in the space.

Marker output will be higher quality than docling output across most doc types, especially with the --use_llm flag. A few specific things we do differently:

  - We have hybrid mode with gemini that merges tables across pages, improves quality on forms, etc.
  - we run an ordering model, so ordering is better for docs where the PDF orde ris bad
  - OCR is a lot better, we train our own model, surya - https://github.com/VikParuchuri/surya
  - References and links
  - Better equation conversion (soon including inline)

Marker (https://www.github.com/VikParuchuri/marker) works kind of like this. It uses a layout model to identify blocks and processes each one separately. The internal format is a tree of blocks, which have arbitrary fields, but can all render to html. It can write out to json, html, or markdown.

I integrated gemini recently to improve accuracy in certain blocks like tables. (get initial text, then pass to gemini to refine) Marker alone works about as well as gemini alone, but together they benchmark much better.

It's on the list to build - been focusing on quality pretty heavily lately.

Hey, I'm the author of marker - thanks for sharing. Most of the processing time is model inference right now. I've been retraining some models lately onto new architectures to improve speed (layout, tables, LaTeX OCR).

We recently integrated gemini flash (via the --use_llm flag), which maybe moves us towards the "hybrid system" you mentioned. Hoping to add support for other APIs soon, but focusing on improving quality/speed now.

Happy to chat if anyone wants to talk about the difficulties of parsing PDFs, or has feedback - email in profile.

Datalab | NYC | Full-time | Software Engineer and Head of Business Ops | $250k-$350k + 1.5-3% equity | https://www.datalab.to

A significant % of useful data is locked away in tough-to-parse formats like PDFs. We build tools to extract it, like https://github.com/VikParuchuri/surya (15k Github stars), and https://github.com/VikParuchuri/marker (19k stars). We also run an inference API and product.

We do meaningful research (we’ve trained several SoTA models), ship product, and contribute to open source. We’re hiring for 2 roles to help us scale:

Senior fullstack software engineer

  - work across our open source repos, inference api, and frontend product
  - interact with our user community
  - you’ll need to be pragmatic, and embrace “boring” technology.  Our stack is fastapi, pytorch, htmx, postgres, and redis.  We deploy to render, and do inference with serverless gpus.
  - requires having built something impressive, ideally an open source project

Head of business operations

  - first non-technical hire
  - work across multiple areas, including finance, hiring, and sales
  - you’ll need to be extremely organized and able to get a lot done
  - requires experience leading operations at an early stage

Email [email protected] if you’re interested - include a link to something you’ve built if possible. You can also read more here - https://datalab-to.notion.site/careers-jan .

Hi, I'm the author of surya (https://github.com/VikParuchuri/surya) - working on improving speed and accuracy now. Happy to collaborate if you have specific page types it's not working on. For modern/clean documents it benchmarks very similarly to Google Cloud, but working on supporting older documents better now.

Hi, I'm the author of marker - https://github.com/VikParuchuri/marker - from my testing, marker handles almost all the issues you mentioned. The biggest issue (that I'm working on fixing right now) is formatting tables properly.

Working on improving tables soon (I'm the author of marker)

This isn't specifically tuned for tables (more for general pdf to markdown), but it's worked for some people with similar use-cases - https://github.com/VikParuchuri/marker

For PDF to markdown, I recently released V2 of my tool marker - https://github.com/vikparuchuri/marker

It should be possible to call a GPL library in a separate process (surya can batch process from the CLI) and avoid GPL - ocrmypdf does this with ghostscript.

I've benchmarked against google cloud ocr, but the results are on Twitter, not the repo yet - https://twitter.com/VikParuchuri/status/1765440195124691339 . The reason I didn't benchmark against doctr is language support.

This looks great! You might be interested in surya - https://github.com/VikParuchuri/surya (I'm the author). It does OCR (much more accurate than tesseract), layout analysis, and text detection.

The OCR is slow on CPU (working on it), but faster than tesseract (CPU-only) on GPU.

You could probably replace pymupdf, tesseract, and some layout heuristics with this.

Happy to discuss more, feel free to email me (in profile).

I unfortunately haven't had time to benchmark against more than tesseract.

You may want to try https://github.com/VikParuchuri/surya (I'm the author). I've only benchmarked against tesseract, but it outperforms it by a lot (benchmarks in repo). Happy to discuss.

You could also try https://github.com/VikParuchuri/marker for general PDF parsing (I'm also the author) - it seems like you're more focused on tables.

Hi HN - I released an open source OCR model yesterday that supports 93 world languages. It builds on a text line detector I created earlier.

In my benchmarks, it's more accurate than tesseract in every language except one. (see repo for benchmarking method)

Since it can run on GPU, speed is about equal to tesseract (when cost-matched with a 1x lambda A6000 vs 28 DigitalOcean CPU cores).

It's built using a modified donut architecture - I added an MoE layer, GQA for faster decoding, and UTF-16 decoding (can represent any character, and faster than UTF-8 since you can combine adjacent bytes.)

I theorized that character-level decoding would be an optimal compute allocation, and that a large embedding matrix (relative to UTF-8 decoding) would store language-specific information.

I trained it using 4x A6000s for about 2 weeks.

You can run surya via Python API, from the CLI, or via an interactive app in the repo.

The size of the framework is not the most important factor - the model weights are usually 10x+ the size of the framework.

The most important factor is inference speed. For something called Nitro, I really expected speed benchmarks. I'd be interested in CPU, CUDA, and MPS at different batch sizes.

Thanks for letting me know. I see the C swapped to an epsilon - the new model checkpoint (live now) seems to fix that. The \, rendered as , is due to Github math rendering (it renders fine elsewhere). I'll manually edit to remove those.

Hi HN - I made texify to convert equations to markdown/LaTeX for my project marker [1] then realized it could be generally useful.

Texify converts equations and surrounding text to Markdown, with embedded LaTeX (MathJax compatible). You can either use a GUI to select equations (inline or block) from PDFs and images to convert, or use the CLI to batch convert images. It works on CPU, GPU, or MPS (Mac).

The closest open source comparisons are pix2tex and nougat - marker is much more accurate than both of them for this task. However, nougat is more for entire pages, and pix2tex is more for block equations (not inline equations and text).

I trained texify for 2 days on 4x A6000 GPUs - I was pleasantly surprised how far I could get with limited GPU resources by reframing the problem to use small parameter counts/images.

Texify is licensed for commercial use, with the weights under CC-BY-SA 4.0. Find them at [2] .

See the texify repo [3] for more details, benchmarks, how to install, etc.

[1] https://github.com/VikParuchuri/marker

[2] https://huggingface.co/vikp/texify

[3] https://github.com/VikParuchuri/texify