Datasets:

nielsr
/

repro-rethinking-genomic-modeling-through-optical-character-recognition-agent-traces

Modalities:

Formats:

Size:

ArXiv:

Tags:

reproducibility

Libraries:

Dataset card Data Studio Files Files and versions

Codex traces preview

Sat, Jul 18

=> can you run the following prompt, but on this version of Trackio please? [gradio-app/trackio#634](https://github.com/gradio-app/trackio/pull/634). We want to test a new version of Trackio # First, read the challenge instructions: curl -sL https://huggingface.co/datasets/ICML-2026-agent-repro/challenge/resolve/main/README.md Your job is to reproduce the ICML 2026 paper #23028 — Rethinking Genomic Modeling Through Optical Character Recognition (OpenReview id: nggzekChuU). Here are the major claims that you should verify: - Claim 1: On the DNALONGBENCH eQTL prediction task, OpticalDNA's linear probe achieves an average AUROC of 0.852 across 9 GTEx tissues, versus 0.514 for HyenaDNA and 0.791 for JanusDNA (DNALONGBENCH eQTL results section). - Claim 2: OpticalDNA exceeds a 985x-larger expert model (AUROC 0.681) on the eQTL task while using only 256K trainable parameters for the linear probe (DNALONGBENCH eQTL results section). - Claim 3: On RiceSubBench, OpticalDNA achieves the best accuracy on all splits, including 0.590 on in-domain japonica and 0.599 on far-OOD glaberrima, beating LucaOne's 0.526 on the far-OOD split by +13.88% (RiceSubBench results section). - Claim 4: On RiceWGPB whole-genome phenotype prediction over ~400M-base genomes, OpticalDNA completes inference in 12.3 minutes, versus 32.5 minutes for LucaOne (1.8B parameters) and 5 hours 40 minutes for Evo-2 (7B parameters) (RiceWGPB results section). - Claim 5: OpticalDNA's visual rendering of DNA achieves a 19.0-21.2x compression ratio from bases to visual tokens, yielding best overall genomic benchmark performance with nearly 20x fewer effective tokens on sequences up to 450k bases (Token Efficiency section). - Claim 6: OpticalDNA is pretrained on 6 OCR-style primitive tasks (T1-T6: free-form transcription, text+localization, ROI transcription, masked completion, subsequence retrieval, and chromosome classification) using a SAM-Conv-CLIP-L visual encoder and a DeepSeek-3B document decoder with 570M activated parameters (Pretraining Tasks / Architecture Components section). # Step 1 — scaffold the canonical logbook (mandatory): curl -sL https://huggingface.co/spaces/ICML-2026-agent-repro/challenge/raw/main/scripts/scaffold_icml_logbook.py | python3 - --title "Rethinking Genomic Modeling Through Optical Character Recognition" --orid "nggzekChuU" --arxiv "2602.02014" --openreview-url "https://openreview.net/forum?id=nggzekChuU" --hf-indexed --claims-json '["On the DNALONGBENCH eQTL prediction task, OpticalDNA'\''s linear probe achieves an average AUROC of 0.852 across 9 GTEx tissues, versus 0.514 for HyenaDNA and 0.791 for JanusDNA (DNALONGBENCH eQTL results section).","OpticalDNA exceeds a 985x-larger expert model (AUROC 0.681) on the eQTL task while using only 256K trainable parameters for the linear probe (DNALONGBENCH eQTL results section).","On RiceSubBench, OpticalDNA achieves the best accuracy on all splits, including 0.590 on in-domain japonica and 0.599 on far-OOD glaberrima, beating LucaOne'\''s 0.526 on the far-OOD split by +13.88% (RiceSubBench results section).","On RiceWGPB whole-genome phenotype prediction over ~400M-base genomes, OpticalDNA completes inference in 12.3 minutes, versus 32.5 minutes for LucaOne (1.8B parameters) and 5 hours 40 minutes for Evo-2 (7B parameters) (RiceWGPB results section).","OpticalDNA'\''s visual rendering of DNA achieves a 19.0-21.2x compression ratio from bases to visual tokens, yielding best overall genomic benchmark performance with nearly 20x fewer effective tokens on sequences up to 450k bases (Token Efficiency section).","OpticalDNA is pretrained on 6 OCR-style primitive tasks (T1-T6: free-form transcription, text+localization, ROI transcription, masked completion, subsequence retrieval, and chromosome classification) using a SAM-Conv-CLIP-L visual encoder and a DeepSeek-3B document decoder with 570M activated parameters (Pretraining Tasks / Architecture Components section)."]' Use Space title `Reproduction: Rethinking Genomic Modeling Through Optical Character Recognition` and publish slug `repro-rethinking-genomic-modeling-through-optical-character-recognition` (never the OpenReview id). Logbook structure (fixed order): - Index: title + paper link + Pages table only - Executive summary: pinned summary + Scope & cost table + pinned Chenruishuo/posterly poster (`poster_embed.html`) - Claim 1 … Claim N: one page per claim - Conclusion: reproduction bundle artifact + download/rerun instructions Try to verify the claims as much as you can locally or using Hugging Face Jobs. Link every Hub model, dataset, Job, Bucket, and GitHub repo in logbook cells. # Last step — validate, then publish: curl -sL https://huggingface.co/spaces/ICML-2026-agent-repro/challenge/raw/main/scripts/validate_icml_logbook.py | python3 - --space /repro-rethinking-genomic-modeling-through-optical-character-recognition && trackio logbook publish /repro-rethinking-genomic-modeling-through-optical-character-recognition Print the published logbook link here when done. 10 messages 16 tools

OpticalDNA reproduction — Codex agent trace

This dataset contains the raw Codex JSONL session trace for the ICML 2026 reproduction of Rethinking Genomic Modeling Through Optical Character Recognition.

The JSONL is uploaded directly from the matching ~/.codex/sessions entry, as recommended by the Agent Trace Viewer. It captures the reproduction work, Hugging Face Jobs audit, poster creation, Trackio rendering fixes, publication, and subsequent logbook QA iterations. The published file is necessarily a point-in-time snapshot: the final action that uploads the trace cannot be contained within the same immutable upload snapshot.

Before publication, non-encrypted trace fields were scanned for token-shaped credentials. No access tokens or API keys were detected. Encrypted Codex reasoning payloads were left unchanged.

Downloads last month: 123

Number of rows:

Total file size:

Paper for nielsr/repro-rethinking-genomic-modeling-through-optical-character-recognition-agent-traces

Rethinking Genomic Modeling Through Optical Character Recognition

Paper • 2602.02014 • Published Feb 2 • 1