feat: code for sft purpcode and baselines

[ganler]

No description provided.

[Copilot]

Pull Request Overview

This PR introduces training configurations and scripts for the SFT (Supervised Fine-Tuning) component of the PurpCode project, including baseline implementations and controlled experiments for data ablation studies.

• Adds training configurations for context distillation fine-tuning with Qwen 14B and 32B models
• Implements data splitting script for safety ratio ablation experiments with different ratios (1/3, 1/2, 2/3)
• Provides baseline training configurations for SafeCoder and ProSec-SIMPO methods
• Includes utility scripts for pushing models and datasets to Hugging Face Hub

Reviewed Changes

Copilot reviewed 10 out of 10 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
sft/ctxdistill_qwen32b.yaml	Configuration for 32B model context distillation training
sft/ctxdistill_qwen14b.yaml	Configuration for 14B model context distillation training
sft/controlled/data_split_for_data_ablation.py	Script to split data for safety ratio ablation experiments
sft/controlled/ctxdistill_qwen14b_safety_ratio_*.yaml	Training configs for different safety data ratios
sft/baselines/safecoder.yaml	SafeCoder baseline training configuration
sft/baselines/prosec-simpo.yaml	ProSec-SIMPO baseline training configuration
script/push_model_hub.py	Fixed example command comment
script/push_data_hub.py	Utility script for pushing datasets to Hugging Face Hub

Comments suppressed due to low confidence (3)

script/push_data_hub.py:9

• The function name 'push_model' is misleading as this function pushes datasets, not models. It should be renamed to 'push_dataset' or 'push_data' for clarity.

def push_model(path: str, split: str, dataset: str = None):