juanako-7b-v1 (UNA: Uniform Neural Alignment)

This model uses uniform neural alignment (UNA) for the DPO training phases and is a fine-tuned version of fblgit/zephyr-lora-dpo-b1 on the HuggingFaceH4/ultrafeedback_binarized dataset.

It is recommended to use the latest Juanako Version which highly outperforms the v1

It achieves the following results on the evaluation set:

• Loss: 0.4594
• Rewards/chosen: -1.1095
• Rewards/rejected: -2.3132
• Rewards/accuracies: 0.7964
• Rewards/margins: 1.2037
• Logps/rejected: -220.0052
• Logps/chosen: -217.5506
• Logits/rejected: -2.5535
• Logits/chosen: -2.7973

Followed alignment-handbook to perform DPO (Phase 2) over Zephyr-SFT model.

Please feel free to run more tests and commit the results. Also if you are interested to participate in UNA's paper research or GPU sponsorship to support UNA research, feel free to contact.

Special thanks to TheBloke for converting the model into multiple formats and overall his enormous contribution to the community. Here are the models:

• juanako-7B-v1-AWQ
• juanako-7B-v1-GPTQ
• juanako-7B-v1-GGUF

Prompt and Inference Usage

# Install transformers from source - only needed for versions <= v4.34
# pip install git+https://github.com/huggingface/transformers.git
# pip install accelerate

import torch
from transformers import pipeline

pipe = pipeline("text-generation", model="fblgit/juanako-7b-v1", torch_dtype=torch.float16, device_map="auto")

# We use the tokenizer's chat template to format each message - see https://huggingface.co/docs/transformers/main/en/chat_templating
messages = [
    {
        "role": "system",
        "content": "You are a friendly chatbot who always responds in the style of a pirate",
    },
    {"role": "user", "content": "How many helicopters can a human eat in one sitting?"},
]
prompt = pipe.tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
outputs = pipe(prompt, max_new_tokens=256, do_sample=True, temperature=0.7, top_k=50, top_p=0.95)
print(outputs[0]["generated_text"])
# <|system|>
# You are a friendly chatbot who always responds in the style of a pirate.</s>
# <|user|>
# How many helicopters can a human eat in one sitting?</s>
# <|assistant|>
# Ah, me hearty matey! But yer question be a puzzler! A human cannot eat a helicopter in one sitting, as helicopters are not edible. They be made of metal, plastic, and other materials, not food!

Model description

It seems to outperforms the original Zephyr in most of the tasks.

I trained Juanako with the same datasets and trainer from alignment-handbook/zephyr-7b-sft-lora

• 1 epoch on DPO with transformers-UNA, the result is fblgit/zephyr-lora-dpo-b1 after merge using FastChat converter.
• finally 1 epoch on DPO with transformers-UNA to fblgit/zephyr-lora-dpo-b1.

Some other experiments were performed as well to test transformers-UNA capabilities on diverse scenarios and models.

This is a complete version of the model, the result of converting LoRa's

Intended uses & limitations

Research purposes.

Training and evaluation data

alignment-handbook DPO with UNA on top of the SFT lora.

Evaluation lm-evaluation-harness

GSM8K 5-Shot

hf (pretrained=fblgit/juanako-7b-v1,load_in_4bit=False,dtype=float16), limit: None, num_fewshot: 5, batch_size: 4

0-Shot Tests

hf (pretrained=fblgit/juanako-7b-v1,load_in_4bit=False,dtype=float16), limit: None, num_fewshot: 0, batch_size: 8

hf (pretrained=fblgit/juanako-7b-v1,load_in_4bit=False,dtype=float16), limit: None, num_fewshot: 25, batch_size: 1

HellaSwag 10-Shot

hf (pretrained=fblgit/juanako-7b-v1,load_in_4bit=False,dtype=float16), limit: None, num_fewshot: 10, batch_size: 16

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0001
• train_batch_size: 1
• eval_batch_size: 1
• seed: 42
• distributed_type: multi-GPU
• num_devices: 12
• gradient_accumulation_steps: 16
• total_train_batch_size: 192
• total_eval_batch_size: 12
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_ratio: 0.01
• num_epochs: 1

Training results

Framework versions

• Transformers 4.35.0-UNA
• Pytorch 2.1.0
• Datasets 2.14.6
• Tokenizers 0.14.1

MMLU Results

5-Shot

hf (pretrained=fblgit/juanako-7b-v1,load_in_4bit=False,dtype=float16), limit: None, num_fewshot: 5, batch_size: 1

Citations

Please feel free to raise a PR if there is any missing citation.

@misc{tunstall2023zephyr, title={Zephyr: Direct Distillation of LM Alignment}, author={Lewis Tunstall and Edward Beeching and Nathan Lambert and Nazneen Rajani and Kashif Rasul and Younes Belkada and Shengyi Huang and Leandro von Werra and Clémentine Fourrier and Nathan Habib and Nathan Sarrazin and Omar Sanseviero and Alexander M. Rush and Thomas Wolf}, year={2023}, eprint={2310.16944}, archivePrefix={arXiv}, primaryClass={cs.LG} } @software{eval-harness, author = {Gao, Leo and Tow, Jonathan and Biderman, Stella and Black, Sid and DiPofi, Anthony and Foster, Charles and Golding, Laurence and Hsu, Jeffrey and McDonell, Kyle and Muennighoff, Niklas and Phang, Jason and Reynolds, Laria and Tang, Eric and Thite, Anish and Wang, Ben and Wang, Kevin and Zou, Andy}, title = {A framework for few-shot language model evaluation}, month = sep, year = 2021, publisher = {Zenodo}, version = {v0.0.1}, doi = {10.5281/zenodo.5371628}, url = {https://doi.org/10.5281/zenodo.5371628} } @misc{rafailov2023direct, title={Direct Preference Optimization: Your Language Model is Secretly a Reward Model}, author={Rafael Rafailov and Archit Sharma and Eric Mitchell and Stefano Ermon and Christopher D. Manning and Chelsea Finn}, year={2023}, eprint={2305.18290}, archivePrefix={arXiv}, }