Recover flattened benchmark tables emitted as paragraph runs

[StevenVincentOne]

Problem

On some born-digital academic PDFs (for example DeepSeek_R1.pdf), benchmark tables are emitted as a run of SemanticParagraph blocks instead of TableBorder, so markdown output loses table structure.

Example symptom: section 3.2 benchmark matrix is flattened into prose-like paragraph lines and no markdown table is emitted.

Repro

• PDF: DeepSeek_R1.pdf
• Parser config: default markdown generation
• Observed: metric header + row values + caption (Table 2 | ...) appear as paragraph sequence rather than table syntax.

Proposal

Add a conservative recovery pass in markdown generation that:

1. Detects a short contiguous paragraph window containing:
   • a metric header pattern (e.g., pass@1, cons@64, rating)
   • numeric row runs with model tokens
   • a Table N | ... caption
2. Reconstructs a markdown table with stable column names.
3. Falls back with no output change unless high-confidence conditions are met.

I have a branch with a first implementation + tests and can open a PR linked to this issue.

[MaximPlusov]

@StevenVincentOne Could you please send us the PDF file you're having trouble with?

[bundolee]

🤖 AI Issue Fix — status: blocked

Analysis: Reporter has a branch with implementation + tests ready. Awaiting:

1. PDF file from reporter (requested by @MaximPlusov)
2. Reporter's PR submission

This is a new recovery feature, not a bug in existing code.

---

Thanks for the detailed proposal and for preparing an implementation! We see you already have a branch with a fix and tests ready. As @MaximPlusov mentioned, could you send us the PDF file so we can validate the approach? Once we have the file and your PR, we'll be happy to review it.

[bundolee]

To send the PDF file, please email it to open.dataloader@hancom.com. This will help us validate the table recovery approach with the actual document.

[StevenVincentOne]

Sorry we didn’t track the specific pdfSent from my iPhoneOn Mar 10, 2026, at 7:44 PM, Bundo Lee ***@***.***> wrote:bundolee left a comment (opendataloader-project/opendataloader-pdf#259) To send the PDF file, please email it to ***@***.*** This will help us validate the table recovery approach with the actual document. —Reply to this email directly, view it on GitHub, or unsubscribe.You are receiving this because you were mentioned.Message ID: ***@***.***>

[MaximPlusov]

Sorry we didn’t track the specific pdf

You wrote about DeepSeek_R1.pdf. Could you send us this file?

[sipa-echo-zaoa]

Hi,
I think @StevenVincentOne is reffering to this pdf DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via
Reinforcement Learning.
It can be downloaded from the link above.

[bundolee]

🤖 AI Issue Fix — status: triaged

Analysis: this issue requires large-scale development beyond automated fix scope.
Label enhancement applied. Requires human implementation.

Update: PDF file now available via arxiv link provided by @sipa-echo-zaoa.
Reporter (@StevenVincentOne) has a branch with implementation + tests ready.
Awaiting PR submission from reporter.

---

Good news — thanks to @sipa-echo-zaoa, the DeepSeek-R1 PDF is now available via the arxiv link. This unblocks validation of the table recovery approach.

@StevenVincentOne, since you mentioned having a branch with implementation and tests ready, would you be able to open a PR? The team can now test against the actual PDF. We're looking forward to reviewing your contribution!

[MaximPlusov]

@bundolee the PR has already been created here: #262.

The document linked above (https://arxiv.org/pdf/2501.12948) doesn’t contain a table in section 3.2, so it’s probably not the one described in the issue.

We found another document DeepSeek_R1.pdf, which includes a table in Section 3.2.