ABForge Data: SFT and RL Training Pools for Ablation Design

A 2026 data release for ABForge, a post-training pipeline for paper-grounded ablation design. The dataset provides Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) training pools and the held-out AblationBench evaluation sets. It was authored by SlowGuess and hosted on Hugging Face.

Use Cases

• Training models for automated ablation design based on the provided SFT trajectory pool.
• Fine-tuning language models for generating target modules and research questions using the task-specific SFT views.
• Benchmarking ablation design algorithms against the held-out AblationBench evaluation sets.
• Conducting reinforcement learning experiments for scientific paper analysis and hypothesis generation.

Strengths

• Includes a raw SFT trajectory pool with 52,813 rows.
• Provides a filtered, task-specific SFT view with 45,961 rows for a concrete ablation design task.
• Contains distinct, held-out evaluation sets (AblationBench) for benchmarking.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count for the full dataset and other specific files is unknown, which may limit suitability assessment.
• Description metadata is limited; actual data quality and structure require manual inspection after download.

Provenance

Source

Hugging Face dataset authored by SlowGuess.

Collection Method

Likely generated as part of the ABForge post-training pipeline for ablation design.

Freshness

Last updated 2026-06-05 07:26:54; freshness should be verified.