tinyllama-moe-base-mix-orpo

<!-- This model card has been generated automatically according to the information the Trainer had access to. You should probably proofread and complete it, then remove this comment. -->

tinyllama-moe-base-mix-orpo

This model is a fine-tuned version of four-two-labs/tinyllama-moe-base on the see code dataset. It achieves the following results on the evaluation set:

• Loss: 1.1757
• Rewards/chosen: -0.1708
• Rewards/rejected: -0.1682
• Rewards/accuracies: 0.5003
• Rewards/margins: -0.0027
• Logps/rejected: -1.6818
• Logps/chosen: -1.7085
• Logits/rejected: -2.6114
• Logits/chosen: -2.6139
• Nll Loss: 1.0859
• Log Odds Ratio: -0.8989
• Log Odds Chosen: -0.1073

Model description

More information needed

Training and evaluation data

from datasets import load_dataset
from datasets import interleave_datasets


def format_chat_template(row):
    for key in ['prompt', 'chosen', 'rejected']: 
        row[key] = tokenizer.apply_chat_template(row[key], tokenize=False)
        
    return row


dataset = (
    interleave_datasets([
        (
            interleave_datasets(
                load_dataset(
                    'four-two-labs/orpo-dpo-mix-40k-multilang-fixed',                 
                    token=hf_token,
                )
                .values()
            )
            .select_columns(['prompt', 'chosen', 'rejected'])
        ),
        (
            load_dataset(
                'four-two-labs/translations-5M-DPO',
                split='train', 
                token=hf_token,                
            )
            .shuffle(42)
            .select(range(250_000))
            .select_columns(['prompt', 'chosen', 'rejected'])
        ),
        (
            load_dataset(
                'four-two-labs/ultrafeedback_binarized-fixed', 
                split='train_prefs', 
                token=hf_token,
            )
            .select_columns(['prompt', 'chosen', 'rejected'])
        ),        
    ])
    .shuffle(seed=42)
    #.select(range(1000))
    .map(format_chat_template, num_proc=32)    
    .train_test_split(test_size=0.01)
)

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 1e-06
• train_batch_size: 12
• eval_batch_size: 12
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 48
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 10
• num_epochs: 3

Training results

Framework versions

• Transformers 4.40.0
• Pytorch 2.1.2+cu121
• Datasets 2.19.0
• Tokenizers 0.19.1