Datasets:

abidlabs
/

repro-memory-savings-at-what-cost-a-study-of-alternatives-to-backpropagation-traces

Modalities:

Formats:

Size:

Tags:

format:agent-traces

Libraries:

Dataset card Data Studio Files Files and versions

Codex traces preview

Wed, Jul 22

# 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 #1499 — Memory Savings at What Cost? A Study of Alternatives to Backpropagation (OpenReview id: QS0ZROPx8u). Here are the major claims that you should verify: - Claim 1: Backpropagation with checkpointing achieves up to 31.1% higher accuracy than FmAD and ZO variants - Claim 2: 34.8% faster convergence and 3.8x fewer computations at comparable memory usage - Claim 3: FmAD and ZO trade memory for higher computational cost and longer wall-clock time to convergence Logbook structure (fixed order): - Index: title + 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: overall findings and reproducibility notes 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-memory-savings-at-what-cost-a-study-of-alternatives-to-backpropagation && trackio logbook publish /repro-memory-savings-at-what-cost-a-study-of-alternatives-to-backpropagation Print the published logbook link here when done. 3 messages 16 tools

Agent traces

Agent sessions published from a Trackio Logbook.

Downloads last month: 50

Number of rows:

Total file size: