DeltaSynth: Experimental Data for Agentic Visual and World Model Evaluation

Companion data release for the 2026 paper 'Δ-Harness: An Agentic Data Harness for Generative Visual and World Models'. The dataset contains experimental data produced and consumed by the DeltaSynth pipeline for LoRA training and held-out evaluation. The code and pipeline are available on GitHub under haolpku/DeltaSynth.

Use Cases

• Training LoRA adapters for generative models based on the experimental data.
• Evaluating the performance of agentic data harnesses on held-out tasks.
• Benchmarking generative visual and world models using the provided evaluation harness.
• Reproducing experiments from the associated 2026 OpenDCAI technical report.

Strengths

• Directly associated with a published 2026 technical report from OpenDCAI.
• Includes a full pipeline, harness, planners, and evaluation code on GitHub, suggesting reproducibility.

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download.
• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.

Provenance

Source

Author lhpku20010120, associated with the GitHub repository haolpku/DeltaSynth.

Collection Method

Produced as experimental data by the DeltaSynth pipeline described in the associated paper.

Freshness

Last updated 2026-06-21 10:12:35; freshness should be verified.