Suzume ORPO

[Paper] [Dataset]

This is Suzume ORPO, an ORPO trained fine-tune of the lightblue/suzume-llama-3-8B-multilingual model using our lightblue/mitsu dataset.

We have trained several versions of this model using ORPO and so recommend that you use the best performing model from our tests, lightblue/suzume-llama-3-8B-multilingual-orpo-borda-half.

Note that this model has a non-commerical license as we used the Command R and Command R+ models to generate our training data for this model (lightblue/mitsu).

We are currently working on a developing a commerically usable model, so stay tuned for that!

Model list

We have ORPO trained the following models using different proportions of the lightblue/mitsu dataset:

• Trained on the top/bottom responses of all prompts in the dataset: lightblue/suzume-llama-3-8B-multilingual-orpo-borda-full
• Trained on the top/bottom responses of the prompts of the 75% most consistently ranked responses in the dataset: lightblue/suzume-llama-3-8B-multilingual-orpo-borda-top75
• Trained on the top/bottom responses of the prompts of the 50% most consistently ranked responses in the dataset: lightblue/suzume-llama-3-8B-multilingual-orpo-borda-half
• Trained on the top/bottom responses of the prompts of the 25% most consistently ranked responses in the dataset: lightblue/suzume-llama-3-8B-multilingual-orpo-borda-top25

Model results

We compare the MT-Bench scores across 6 languages for our 4 ORPO trained models, as well as some baselines:

• meta-llama/Meta-Llama-3-8B-Instruct - The foundation model that our models are ultimately built upon
• Nexusflow/Starling-LM-7B-beta - The highest performing open model on the Chatbot arena that is of a similar size to ours
• gpt-3.5-turbo - A fairly high quality (although not state-of-the-art) proprietary LLM
• lightblue/suzume-llama-3-8B-multilingual - The base model which we train our ORPO finetunes from

We can see noticable improvement on most languages compared to the base model. We also find that our ORPO models achieve the highest score out of all the models we evaluated for a number of languages.

Training data

We trained this model using the lightblue/mitsu_full_borda dataset.

Training configuration

axolotl version: 0.4.0

base_model: lightblue/suzume-llama-3-8B-multilingual
model_type: LlamaForCausalLM
tokenizer_type: AutoTokenizer  # PreTrainedTokenizerFast

load_in_8bit: false
load_in_4bit: false
strict: false

rl: orpo
orpo_alpha: 0.1
remove_unused_columns: false

chat_template: chatml
datasets:
  - path: lightblue/mitsu_tophalf_borda
    type: orpo.chat_template
    conversation: llama-3
dataset_prepared_path: /workspace/llm_training/axolotl/llama3-multilingual-orpo/prepared_mitsu_half_borda
val_set_size: 0.02
output_dir: /workspace/llm_training/axolotl/llama3-multilingual-orpo/output_mitsu_half_borda

sequence_len: 8192
sample_packing: false
pad_to_sequence_len: true

use_wandb: true
wandb_project: axolotl
wandb_entity: peterd
wandb_name: mitsu_half_borda

gradient_accumulation_steps: 8
micro_batch_size: 1
num_epochs: 1
optimizer: paged_adamw_8bit
lr_scheduler: cosine
learning_rate: 8e-6

train_on_inputs: false
group_by_length: false
bf16: auto
fp16:
tf32: false

gradient_checkpointing: true
gradient_checkpointing_kwargs:
  use_reentrant: false
early_stopping_patience:
resume_from_checkpoint:
logging_steps: 1
xformers_attention:
flash_attention: true

warmup_steps: 10
evals_per_epoch: 20
eval_table_size:
saves_per_epoch: 1
debug:
deepspeed: /workspace/axolotl/deepspeed_configs/zero3_bf16.json
weight_decay: 0.0
special_tokens:
  pad_token: <|end_of_text|>

workspace/llm_training/axolotl/llama3-multilingual-orpo/output_mitsu_half_borda

This model is a fine-tuned version of lightblue/suzume-llama-3-8B-multilingual on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.0935

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 8e-06
• train_batch_size: 1
• eval_batch_size: 1
• seed: 42
• distributed_type: multi-GPU
• num_devices: 4
• gradient_accumulation_steps: 8
• total_train_batch_size: 32
• total_eval_batch_size: 4
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 10
• num_epochs: 1

Training results

Framework versions

• Transformers 4.38.2
• Pytorch 2.2.1+cu121
• Datasets 2.18.0
• Tokenizers 0.15.0

How to cite

@article{devine2024sure,
  title={Are You Sure? Rank Them Again: Repeated Ranking For Better Preference Datasets},
  author={Devine, Peter},
  journal={arXiv preprint arXiv:2405.18952},
  year={2024}
}

Developer

Peter Devine - (ptrdvn)