Training
API reference for training scripts and trainers.
SFT Training
CPT Training
Data Processing
Build a mixed Arrow dataset from multiple source datasets with configurable sampling ratios.
- Usage:
python data/build_mixed_arrow_dataset.py –config configs/sft_llada.yaml
The script: 1. Loads each Arrow dataset from config.dataset.arrow_dirs 2. Applies sampling (downsample/upsample) based on config.dataset.sampling_ratios 3. Concatenates all sampled datasets 4. Shuffles globally with the seed from config.training.seed 5. Saves to a single output directory (config.experiment.output_dir/mixed_arrow) 6. Saves metadata (mixing_meta.json) with statistics
- Requirements:
datasets (Hugging Face)
omegaconf
- data.build_mixed_arrow_dataset.sample_dataset(dataset: Dataset, target_ratio: float, seed: int) Dataset[source]
Adjust dataset size by downsampling or upsampling.
- Parameters:
dataset – HuggingFace Dataset
target_ratio – float (e.g., 0.5, 2.0, 1.0) - 1.0: use all samples - < 1.0: downsample without replacement - > 1.0: upsample with replacement (duplicate)
seed – random seed for reproducibility
- Returns:
Dataset with adjusted size
- data.build_mixed_arrow_dataset.build_mixed_dataset(dataset_specs: List[Tuple[str, float]], seed: int, output_dir: str) str[source]
Load, sample, concatenate, shuffle, and save mixed dataset.
- Parameters:
dataset_specs – List of (arrow_dir, ratio) tuples
seed – random seed for reproducibility
output_dir – where to save the mixed dataset
- Returns:
Path to the saved mixed Arrow dataset
Split wildguardmix.jsonl into four safety-categorized files based on harm labels.
Classification (4 groups): - Group A (safe_prompts_safe_response):
prompt_harm_label == “unharmful” AND response_harm_label == “unharmful”
- Group B (safe_prompts_harmful_response):
prompt_harm_label == “unharmful” AND response_harm_label == “harmful”
- Group C (harmful_prompt_safe_response):
prompt_harm_label == “harmful” AND response_harm_label == “unharmful”
- Group D (harmful_prompt_harmful_response):
prompt_harm_label == “harmful” AND response_harm_label == “harmful”
Samples with unknown/missing labels are discarded. All original fields including metadata are preserved in the output files.
- data.split_wildguardmix.classify_sample(item: Dict[str, Any]) str[source]
Classify a sample into one of four groups.
- Returns:
“A” for safe_prompts_safe_response, “B” for safe_prompts_harmful_response, “C” for harmful_prompt_safe_response, “D” for harmful_prompt_harmful_response, None if unknown/invalid