How does this skill help prevent data corruption?

It enforces a 'Fail Fast' policy with multi-step validation, checking magic bytes, header integrity, and version compatibility before attempting to parse tensor data.

Does it support backward compatibility for model files?

Yes, the skill includes specific guidelines for versioning the .mbt format, including strategies for handling breaking changes and implementing legacy loading paths.

Why is I/O logic separated from core layers in this design?

Separating I/O ensures that core tensor and neural network modules remain agnostic of file formats, simplifying maintenance and allowing for the addition of new formats without modifying the core API.

What formats does this skill support for MbTorch?

It provides specifications and implementation patterns for ONNX (importing), safetensors (loading/saving weights), and the native .mbt format (full model serialization).

MbTorch I/O & Model Formats

Name: MbTorch I/O & Model Formats
Author: c-tomioka

byc-tomioka

0•

Data Science & ML

Manages high-performance model I/O for the MbTorch framework, supporting ONNX, safetensors, and native binary formats with strict validation.

This skill transforms Claude into an I/O compatibility engineer for the MbTorch AI/ML framework in MoonBit. It provides comprehensive guidance on implementing robust loaders and writers for ONNX, safetensors, and native .mbt files. By enforcing strict separation of concerns between I/O logic and core tensor operations, the skill ensures that model imports from external frameworks like PyTorch remain reliable. It includes detailed protocols for magic byte validation, versioning strategies, and inference parity testing to prevent silent data corruption and maintain backward compatibility.

Key Features

01Guidance for cross-framework inference parity and roundtrip testing

02Standardized error handling for corrupted data and unsupported operations

03Design patterns for native .mbt binary format evolution and versioning

04Comprehensive validation protocols for magic bytes, headers, and version compatibility

050 GitHub stars

06Strict separation of format-specific logic from core neural network layers

Use Cases

01Developing and versioning a native binary serialization format for edge deployment

02Implementing ONNX model loaders to bridge external ML ecosystems with MoonBit

03Importing pre-trained weights from HuggingFace using the safetensors format

What are Skills?·How to Install

Install with 🐟 Skill.Fish

npx skillfish add c-tomioka/mbtorch mbtorch-io-formats

For use in Claude.ai and ChatGPT

Key Features

01Guidance for cross-framework inference parity and roundtrip testing

02Standardized error handling for corrupted data and unsupported operations

03Design patterns for native .mbt binary format evolution and versioning

04Comprehensive validation protocols for magic bytes, headers, and version compatibility

050 GitHub stars

06Strict separation of format-specific logic from core neural network layers

Use Cases

01Developing and versioning a native binary serialization format for edge deployment

02Implementing ONNX model loaders to bridge external ML ecosystems with MoonBit

03Importing pre-trained weights from HuggingFace using the safetensors format