marco-costa-ml/balatro-imitation-learning

Balatro Imitation Learning Dataset

A large-scale gameplay dataset for imitation learning, state modeling, and downstream vision research in Balatro.

This release contains 494.38 hours of gameplay video, processed into structured detection, OCR, and game-state representations derived from publicly available gameplay footage.

Highlights

• 494.38 hours of gameplay video
• Structured multimodal state extracted from gameplay footage
• Includes cleaned detections, cleaned OCR, tracked object state, stable OCR, and composed object state
• Designed for imitation learning, state reconstruction, and game understanding tasks

Dataset Explorer

Explore the dataset interactively with:

marco-costa-ml/balatro-state-explorer

This is the easiest way to inspect the schema, understand the state representations, and browse examples across the dataset.

Included Data

The dataset is organized as follows:

├── clean
│   ├── det
│   └── rec
├── derived
│   ├── composed_state
│   ├── object_state
│   └── ocr_stable

clean/

Intermediate cleaned outputs from the vision models.

• clean/det/: cleaned object detections
• clean/rec/: cleaned OCR outputs

derived/

Higher-level state representations built from the cleaned model outputs.

• derived/object_state/: temporally consistent object tracks
• derived/ocr_stable/: stabilized OCR time series
• derived/composed_state/: parent-child composed object state with attached attributes

Excluded Data

Some internal pipeline stages are intentionally not included in this release:

• raw/: excluded because the raw model outputs are too large to distribute practically
• derived/event_state/: excluded because this representation is still a work in progress and not yet stable enough for release

So while the internal pipeline contains the following broader structure:

├── clean
│   ├── det
│   └── rec
├── derived
│   ├── composed_state
│   ├── event_state
│   ├── object_state
│   └── ocr_stable
└── raw
    ├── det
    └── rec

this dataset release only includes the cleaned and derived components that are currently most useful for downstream work.

Processing Pipeline

The released data comes from a multi-stage computer vision pipeline:

raw detections / OCR
    -> cleaned detections / OCR
    -> tracked object state + stable OCR
    -> composed object state

The main scripts used in this pipeline are:

clean_to_object_state_det.py
clean_to_stable_ocr.py
object_state_to_composed.py
raw_to_clean_det.py
raw_to_clean_ocr.py

1. raw_to_clean_det.py

Converts raw detection CSV outputs into a validated, standardized parquet dataset.

What it does:

• removes low-confidence detections (conf <= 0.3)
• filters invalid bounding boxes
• removes tiny boxes likely to be noise
• rounds confidence values
• converts bounding box coordinates to integers
• writes cleaned results to det_clean.parquet

Outcome: A clean detection dataset containing only valid, meaningful bounding boxes.

2. raw_to_clean_ocr.py

Converts raw OCR CSV outputs into a cleaned, normalized text dataset.

What it does:

• loads rec_raw.csv
• normalizes text to lowercase
• removes spaces and selected special characters
• keeps both the original text (text_raw) and cleaned text (text_clean)
• skips malformed and empty results
• writes cleaned results to rec_clean.parquet

Outcome: A normalized OCR dataset ready for matching, aggregation, and temporal stabilization.

3. clean_to_object_state_det.py

Converts per-frame detections into temporally consistent tracked objects.

What it does:

• reads cleaned detections from det_clean.parquet
• groups detections by frame
• tracks objects over time independently by class
• matches detections using spatial distance, overlap, and simple motion prediction
• allows short gaps and interpolates where appropriate
• filters low-quality tracks
• writes object_state.parquet

Outcome: A structured object-state dataset where objects have stable identities over time.

4. clean_to_stable_ocr.py

Transforms cleaned OCR into a temporally stable text signal.

What it does:

• applies label-specific regex validation rules
• nulls invalid OCR values
• fills missing values for selected labels using carry-forward logic
• applies label-specific default values
• smooths OCR over time with sliding-window mode filtering
• writes rec_stable.parquet

Outcome: A robust OCR time series with reduced flicker, fewer invalid readings, and improved temporal consistency.

5. object_state_to_composed.py

Builds a higher-level composed object representation by attaching child attributes to parent objects over time.

High level: This script takes tracked objects plus cleaned detections and produces a temporally stable object representation where attached properties are merged into their parent objects.

What it does:

1. matches tracked objects to detections and attaches confidence scores
2. identifies valid parent-child relationships per frame using spatial overlap
3. builds temporal consistency by grouping matches across frame intervals
4. filters noisy relationships using duration, confidence, and gap thresholds
5. resolves conflicts when multiple candidate children compete for the same parent
6. merges child attributes into parent rows and removes absorbed child rows

Examples of attached child attributes include:

• edition
• sticker
• modifier
• seal
• debuffed state

Outcome: A cleaner and more semantically useful object-state representation where attributes are attached directly to the relevant game objects.

In one line: object_state_to_composed.py converts raw detections plus tracking into a clean, temporally stable representation of objects with attached attributes.

Representation Summary

The dataset includes multiple levels of representation, each useful for different tasks.

Clean detection state

Frame-level detections after validation and cleaning.

Useful for:

• object detection analysis
• custom tracking
• low-level vision debugging

Object state

Tracked objects with stable identities across time.

Useful for:

• state reconstruction
• temporal modeling
• imitation learning pipelines that need persistent entities

Stable OCR

Temporally smoothed OCR signals.

Useful for:

• score/state extraction
• text-conditioned modeling
• robust frame-to-frame metadata parsing

Composed state

Parent objects with attached child attributes merged into a denoised higher-level representation.

Useful for:

• downstream state modeling
• semantic game-state understanding
• imitation learning with cleaner entity structure

Models

The dataset was generated using fine-tuned vision models released separately:

• Detection model: marco-costa-ml/balatro-yolo
• OCR model: marco-costa-ml/balatro-ocr

Data Collection

Data was generated from publicly available gameplay streams using a custom computer vision pipeline.

The released dataset does not include raw video or audio.

Notes

This dataset is still evolving, and future releases may introduce schema changes or additional derived representations.

Users should expect:

• possible formatting changes
• occasional inconsistencies in early pipeline outputs
• additional derived views in future releases

Suggested Uses

This dataset is intended for research and engineering work involving:

• imitation learning
• game-state reconstruction
• multimodal temporal modeling
• object tracking in gameplay footage
• OCR-assisted state extraction
• representation learning for structured game environments