You need to agree to share your Hugging Face information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

HOI Planner and Critic training data

Private research dataset for the HOI Planner and Critic experiments. UTF-8 JSONL files have one JSON object per line. Code and training scripts: esther585/hoi-planner-critic.

Planner

  • hoi_v1/train.jsonl: 3,852 training instructions.
  • hoi_v1/complex.jsonl: 768 added complex-action prompts, included in training data.
  • hoi_v1/planner.jsonl: 3,852 GPT teacher plans with generation_prompt and interaction_graph.
  • hoi_v1/planner_sft.jsonl: 3,852 ready-to-train SFT rows with system/user/assistant messages. Use this for Planner training; it requires no images.
  • hoi_v1/test.jsonl and hoi_v1/labels.jsonl: 192 held-out evaluation prompts and annotations, split evenly between simple and complex tasks. Do not train on these.

Critic

  • critic/train.jsonl, critic/val.jsonl, critic/test.jsonl: 3,376 / 164 / 139 image-conditioned SFT rows. Train on train and validate on val.
  • critic/train_unique.jsonl: 2,779-row unbalanced alternative to the decision-balanced main train split.
  • critic/system.txt: exact system prompt used for the included Critic LoRA.
  • critic/images/: all 3,082 unique generated PNGs referenced by the JSONL rows. Each images path is relative to the critic/ directory.
  • critic/missing_images.tar: redundant backup for one PNG that a previous ModelScope upload did not expose as a standalone file. This Hugging Face copy also includes the standalone PNG.

The Critic rows contain messages, images, and metadata. The assistant response selects accept, edit, or regenerate and gives errors, a repair route, and semantic/visual scores. The balanced train split contains 1,688 accept, 844 edit, and 844 regenerate examples. These rows were derived from 3,104 image-generation trajectories and filtered by trajectory quality.

The Critic adapter was trained on the Critic training split. To use the JSONLs with ms-swift, resolve relative image paths to absolute paths after download; the linked code repository provides code/paths.py and code/train_critic.sh for this. No public redistribution license is granted.

Downloads last month
22