anonymous-submission-221/OTA-76k

---
license: cc-by-4.0
task_categories:
- visual-question-answering
language:
- en
tags:
- agent
size_categories:
- 10K<n<100K
---
# Dataset Card for OTA-76k (POIROT Framework)

## Dataset Description
* **Repository:** [anonymous-submission-221/POIROT](https://github.com/anonymous-submission-221/POIROT)

### Dataset Summary

OTA-76k is a large-scale, bounding-box-grounded, multi-step video reasoning dataset designed to train Multimodal Large Language Models (MLLMs) for fine-grained, spatio-temporal deduction. This dataset addresses the common pitfalls of existing models in video reasoning, such as their over-reliance on frame-level perception and outcome-oriented sparse rewards, which often lead to visual noise interference and logical hallucinations.
Built upon the innovative **Observe-Think-Action (O-T-A)** hierarchical reasoning architecture, this dataset facilitates object-level clue discovery and entity tracking by constructing transparent Visualized Chain-of-Thought (V-CoT) traces, breaking through the perceptual limitations of discrete frame sampling.

### Data Splits

The OTA-76k dataset comprises a total of 76,000 high-quality interaction trajectories derived from 20,000 high-quality video clips. The dataset is strictly partitioned into two progressive subsets:
* **SFT (Supervised Fine-Tuning):** 42k trajectories designed for multi-stage fine-tuning (format alignment, basic spatial anchoring, and multi-turn trajectory tuning).
* **RL (Reinforcement Learning):** 34k trajectories tailored for policy optimization using the Spatial-Grounded GDPO (SG-GDPO) framework.