ryen12/sft-cs336-assign5-datasets
SFT CS336 Assignment 5 Datasets This repository contains all training and evaluation datasets for Supervised Fine-Tuning (SFT) experiments (based on CS-336 Assignment-5) Repository Structure Folder Description sft-instruct/ Instruction-tuning datasets sft-reason/ Reasoning SFT datasets SFT Instruction finetuning Datasets for instruction-following SFT is based on UltraChat-200K + SafetyLlama. File Size Description… See the full description on the dataset page: https://huggingface.co/datasets/ryen12/sft-cs336-assign5-datasets.
SFT CS336 Assignment 5 Datasets
This repository contains all training and evaluation datasets for Supervised Fine-Tuning (SFT) experiments (based on CS-336 Assignment-5)
Repository Structure
SFT Instruction finetuning
Datasets for instruction-following SFT is based on UltraChat-200K + SafetyLlama. | File | Size | Description | |------|------|-------------| | sft-instruct/train.jsonl | ~200K examples | Full training dataset | | sft-instruct/test.jsonl | ~20K | Validation dataset for intermediate evaluation | | sft-instruct/sample_train.jsonl | 1K examples | Debug subsample |
Note: You can also find the eval data and configs for: GSM8K, MMLU, Simple Safety Tests, AlpacaEval in sft-instruct/eval folder. See Readme.md for more details.
Trained Checkpoints using this dataset
SFT Reasoning finetuning
Datasets for math reasoning SFT based on hiyouga/math12k.
See Readme.md for more details.
Trained Checkpoints using this dataset
- qwen-2.5-math-sft-all
- qwen-2.5-math-sft-filtered
- qwen-2.5-math-sft-filtered-res-len
- qwen-2.5-math-sft-filtered-2epoch
For training code and experiment details, see: building-from-scratch/sft