What data formats does funsloth-check support?

The skill automatically detects and validates several popular formats including Alpaca, ShareGPT, ChatML, and raw text formats.

How does it help with Chinchilla optimality?

By analyzing your dataset size against your target model and LoRA rank, it calculates a Chinchilla fraction to let you know if your dataset is too small, just right, or potentially redundant.

Can I use local datasets with this skill?

Yes, you can provide either a Hugging Face dataset ID or a local file path such as a .jsonl file for analysis.

What happens after the validation is complete?

The skill generates a structured summary that can be passed directly to the funsloth-train skill to begin the fine-tuning process with verified parameters.

Does it provide recommendations for training parameters?

Yes, based on the analysis, it suggests sequence length adjustments, learning rates for small datasets, and specific standardization steps for ShareGPT data.

Funsloth Dataset Validator

Name: Funsloth Dataset Validator
Author: chrisvoncsefalvay

bychrisvoncsefalvay

•

データサイエンスとML

Validates and analyzes datasets to ensure they are optimized for Unsloth fine-tuning and Chinchilla optimality.

Funsloth Dataset Validator is a specialized skill designed to streamline the preparation phase of LLM fine-tuning. It automatically detects common data formats such as Alpaca, ShareGPT, and ChatML, while performing rigorous schema validation to catch errors before training begins. The skill provides deep insights through token analysis—identifying sequences that are too short or too long—and calculates Chinchilla optimality fractions to help developers determine if their dataset size is appropriate for their target model and LoRA configuration.

主な機能

01Automatic format detection for Alpaca, ShareGPT, ChatML, and Raw text

02Chinchilla optimality calculations for specific model sizes and LoRA ranks

03Seamless context handoff to training workflows like funsloth-train

044 GitHub stars

05Comprehensive token statistics including mean, median, and outlier detection

06Schema validation with actionable suggestions for data fixes

ユースケース

01Determining the ideal learning rate and sequence length based on dataset metrics

02Verifying a Hugging Face dataset's compatibility with Llama 3.1 training

03Analyzing token distribution to prune problematic sequences before fine-tuning

What are Skills?·How to Install

Install with 🐟 Skill.Fish

npx skillfish add chrisvoncsefalvay/funsloth funsloth-check

For use in Claude.ai and ChatGPT

Download Skill