Caesarrr/interleaved-umm

Interleaved Multimodal Reasoning Dataset

A dataset generation framework for spatial reasoning tasks involving camera viewpoint prediction and ordering around static 3D objects. This project generates multimodal chain-of-thought reasoning traces that teach models how camera views change during orbital rotation.

Overview

This framework generates two types of spatial reasoning tasks:

1. Task 1: Camera View Prediction - Given an initial view and rotation parameters (angle + direction), predict what the object looks like from the new viewpoint
2. Task 3: Camera View Ordering - Given a reference frame and scrambled candidate images, reconstruct the correct temporal order of camera views

Key Features

• Automatic Ground Plane Estimation: PCA-based geometry calibration eliminates manual tuning
• Oracle Chain Generation: Creates step-by-step reasoning paths with intermediate ground-truth views
• LLM Chain-of-Thought: Generates natural language reasoning that mirrors human spatial thinking
• Multi-backend Support: Works with OpenAI-compatible APIs and local vLLM inference
• Cluster Deployment: Ready for distributed GPU execution via Determined AI

Installation

Prerequisites

• Python 3.12
• CUDA 11.8+ (for GPU support)
• Access to CO3D dataset
• (Optional) Determined AI cluster for distributed training

Setup

1. Clone the repository

git clone <repository-url>
cd interleaved-umm

2. Create conda environment

conda create -n interleaved-umm python=3.12
conda activate interleaved-umm

3. Install PyTorch

First, install PyTorch 2.8.0 matching your CUDA version from the official PyTorch website.

For example, with CUDA 11.8:

pip install torch==2.8.0 torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118

For CUDA 12.1:

pip install torch==2.8.0 torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121

4. Install in editable mode

pip install -e .

5. Install dependencies

pip install -r requirements.txt

6. Set up environment variables

Create a .env file in the project root:

# OpenAI-compatible API
BASE_URL=https://api.openai.com/v1/chat/completions
API_KEY=your_api_key_here

# Qwen API (optional)
QWEN_API_KEY=your_qwen_key
QWEN_BASE_URL=https://dashscope.aliyuncs.com/compatible-mode/v1

⚠️ Important: Never commit the .env file. It's already in .gitignore.

7. Prepare CO3D dataset

Download and extract the CO3D dataset, then update paths in generation scripts:

• ROOT_PATH: Path to CO3D dataset root
• IMAGE_PREFIX: Relative path for storing processed images

Project Structure

interleaved-umm/
├── src/
│   ├── action_state/          # Task generation logic
│   │   ├── gen_task1.py       # Camera view prediction
│   │   ├── gen_task3.py       # Camera view ordering
│   │   └── utils.py           # Geometry utilities (PCA, camera poses)
│   ├── llm_generation/        # Chain-of-thought generation
│   │   ├── generator.py       # CoTGenerator orchestrator
│   │   ├── prompts.py         # Task-specific prompts
│   │   ├── api_client.py      # OpenAI-compatible API client
│   │   ├── vllm_client.py     # Local vLLM inference
│   │   └── cleaning_generator.py  # Data quality verification
│   └── utils/
│       └── image_utils.py     # Multimodal content parsing
├── scripts/
│   ├── action_state/          # Task generation runners
│   │   ├── task1/            # Task 1 generation scripts
│   │   └── task3/            # Task 3 generation scripts
│   ├── run_llm_cot.py        # LLM CoT generation (API)
│   ├── run_llm_cot_vllm.py   # LLM CoT generation (vLLM)
│   ├── run_cleaning.py       # Data quality checker
│   ├── filter/               # Sequence filtering scripts
│   ├── copy_image.py         # Image preprocessing
│   └── visualize_*.py        # Visualization tools
├── deploy/
│   ├── local/                # Cluster deployment configs
│   │   ├── task1/
│   │   ├── task3/
│   │   └── cleaning/
│   └── template/             # Config templates
├── configs/                  # Legacy configuration files
├── data/                     # Generated datasets (not in repo)
├── debug/                    # Debugging outputs (not in repo)
├── pyproject.toml           # Package configuration
├── requirements.txt         # Python dependencies
└── CLAUDE.md               # Documentation for Claude Code

Quick Start

1. Generate Task Metadata

Generate Task 1 samples with oracle reasoning chains:

cd scripts/action_state/task1
bash run_gen_task1_v3.sh

This will:

• Sample camera pose pairs from CO3D sequences
• Verify geometric constraints (angle ranges, intervals)
• Generate oracle chains with intermediate views
• Save JSONL files to data/questions/task1_metadata_v3/

2. Generate Chain-of-Thought Reasoning

Option A: Using OpenAI-compatible API

python scripts/run_llm_cot.py \
    --input_file data/questions/task1_metadata_v3/train/train_1.jsonl \
    --output_file data/questions/task1_v3/train/train_1.jsonl \
    --image_root /path/to/project/root \
    --model gpt-4o

Option B: Using local vLLM server

python scripts/run_llm_cot_vllm.py \
    --input_file data/questions/task1_metadata_v3/train/train_1.jsonl \
    --output_file data/questions/task1_v3/train/train_1.jsonl \
    --image_root /path/to/project/root \
    --model /path/to/Qwen3-VL-32B-Instruct \
    --tp_size 2 \
    --gpu_memory_utilization 0.9

3. Deploy to Cluster

If using Determined AI:

det experiment create deploy/local/task1/config.yaml .

Usage

Task Generation Parameters

Task 1 (Camera View Prediction)

Key parameters in scripts/action_state/task1/run_gen_task1_v3.sh:

MIN_ANGLE=60.0        # Minimum rotation angle (degrees)
MAX_ANGLE=125.0       # Maximum rotation angle (degrees)
MIN_INTERVAL=25.0     # Minimum angular separation between options
NUM_SAMPLES=3         # Samples per sequence

Task 3 (Camera View Ordering)

Key parameters in scripts/action_state/task3/run_gen_task3_v3.sh:

MIN_INTERVAL=15.0     # Minimum per-step rotation
MAX_INTERVAL=40.0     # Maximum per-step rotation
MAX_ANGLE=170.0       # Maximum total trajectory span

Data Filtering

Before generating tasks, filter sequences for quality:

python scripts/filter/filter_v4.py \
    --category apple \
    --root_path /path/to/co3d \
    --output_dir data/filter_log_v4_pca

Visualization

Visualize camera trajectories:

python scripts/visualize_traj_pca.py \
    --category apple \
    --root_path /path/to/co3d \
    --sequence_name <sequence_id>

Configuration

Environment Variables

Variable	Description	Example
BASE_URL	OpenAI-compatible API endpoint	https://api.openai.com/v1/chat/completions
API_KEY	API authentication key	sk-...
QWEN_API_KEY	Qwen API key (optional)	sk-...
QWEN_BASE_URL	Qwen API endpoint (optional)	https://dashscope.aliyuncs.com/...

Cluster Deployment

Edit deploy/local/*/config.yaml:

resources:
  resource_pool: amp-80g    # GPU pool
  slots_per_trial: 2        # Number of GPUs

bind_mounts:
  - host_path: /mount/HOME/username
    container_path: /home/username

environment:
  image: your-docker-image:tag

Development

Running Tests

# Test on a small subset
python src/action_state/gen_task1.py \
    --root_path /path/to/co3d \
    --output_dir test_output \
    --category apple \
    --num_samples 1

Code Structure

Geometry Pipeline:

1. CO3DDataLoader loads frame annotations
2. get_sequence_geometry_pca() estimates ground plane via PCA
3. get_relative_yaw() computes angular differences
4. decompose_angle() breaks rotations into steps

CoT Generation Pipeline:

1. CoTGenerator receives oracle chain
2. For each step, constructs context messages
3. Calls LLM with "cheat sheet" (target view + physics hints)
4. LLM generates reasoning that appears to derive the action
5. Combines into final <think>...</think> trace

Key Concepts

• Oracle Chain: Ground-truth reasoning path with intermediate views
• Cheat Mechanism: LLM sees target but must write as if deriving it
• Parallax Rule: "Camera moves RIGHT → View shifts LEFT"
• Bird's Eye View: Rotation direction defined from top-down perspective

Troubleshooting

Issue: FileNotFoundError for images

• Solution: Check IMAGE_PREFIX and image_root match your actual paths

Issue: LinAlgError in PCA

• Solution: Sequence has too few frames or degenerate geometry. Filter will catch these.

Issue: vLLM OOM errors

• Solution: Reduce gpu_memory_utilization or limit_mm_per_prompt

Issue: No valid samples generated

• Solution: Relax MIN_ANGLE, MAX_ANGLE, or MIN_INTERVAL constraints

Citation

If you use this dataset or codebase, please cite:

@misc{interleaved-umm,
  title={Interleaved Multimodal Reasoning Dataset},
  author={Your Name},
  year={2024}
}

License

[Specify your license here]

Contact

For questions or issues, please contact [your contact info] or open an issue on GitHub.