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

Tasks	Version	Filter	Metric	Value		Stderr
gsm8k	Yaml	get-answer	exact_match	0.4761	±	0.0138

0-Shot Tests

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

Tasks	Version	Filter	Metric	Value		Stderr
arc_challenge	Yaml	none	acc	0.5691	±	0.0145
		none	acc_norm	0.6041	±	0.0143
arc_easy	Yaml	none	acc	0.8363	±	0.0076
		none	acc_norm	0.8161	±	0.0079
hellaswag	Yaml	none	acc	0.6554	±	0.0047
		none	acc_norm	0.8411	±	0.0036
boolq	Yaml	none	acc	0.8355	±	0.0065
lambada	N/A	none	perplexity	3.3607	±	0.1398
		none	acc	0.7309	±	0.0137
piqa	Yaml	none	acc	0.8194	±	0.0090
		none	acc_norm	0.8335	±	0.0087
sciq	Yaml	none	acc	0.9480	±	0.0070
		none	acc_norm	0.8960	±	0.0097
truthfulqa	N/A	none	bleu_max	26.0803	±	0.6528
- truthfulqa_mc1	Yaml	none	acc	0.4198	±	0.0173
- truthfulqa_mc2	Yaml	none	acc	0.5847	±	0.0153
winogrande	Yaml	none	acc	0.7609	±	0.0120

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

Tasks	Version	Filter	Metric	Value		Stderr
arc_challenge	Yaml	none	acc	0.6058	±	0.0143
		none	acc_norm	0.6485	±	0.0140

HellaSwag 10-Shot

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

Tasks	Version	Filter	Metric	Value		Stderr
hellaswag	Yaml	none	acc	0.6582	±	0.0047
		none	acc_norm	0.8513	±	0.0036

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

Training Loss	Epoch	Step	Validation Loss	Rewards/chosen	Rewards/rejected	Rewards/accuracies	Rewards/margins	Logps/rejected	Logps/chosen	Logits/rejected	Logits/chosen
0.4966	0.15	50	0.4893	-1.1759	-2.2914	0.7485	1.1155	-219.7872	-218.2148	-2.5450	-2.7884
0.4522	0.31	100	0.4808	-0.8099	-1.8893	0.7784	1.0794	-215.7659	-214.5544	-2.5644	-2.8095
0.5048	0.46	150	0.4706	-1.0526	-2.1412	0.7725	1.0887	-218.2852	-216.9814	-2.5638	-2.8089
0.4853	0.62	200	0.4640	-1.0787	-2.2821	0.7725	1.2034	-219.6941	-217.2426	-2.5460	-2.7891
0.4639	0.77	250	0.4636	-1.2348	-2.4583	0.8084	1.2235	-221.4559	-218.8034	-2.5533	-2.7970
0.4634	0.93	300	0.4601	-1.1370	-2.3243	0.7964	1.1873	-220.1163	-217.8257	-2.5540	-2.7977
-	1.00	300	0.4594	-1.1095	-2.3132	0.7964	1.2037	-220.0052	-217.5506	-2.5535	-2.7973

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

Tasks	Version	Filter	Metric	Value		Stderr
mmlu	N/A	none	acc	0.6236	±	0.1269
- humanities	N/A	none	acc	0.5651	±	0.1289
- formal_logic	Yaml	none	acc	0.4365	±	0.0444
- high_school_european_history	Yaml	none	acc	0.7636	±	0.0332
- high_school_us_history	Yaml	none	acc	0.8039	±	0.0279
- high_school_world_history	Yaml	none	acc	0.7848	±	0.0268
- international_law	Yaml	none	acc	0.7686	±	0.0385
- jurisprudence	Yaml	none	acc	0.7778	±	0.0402
- logical_fallacies	Yaml	none	acc	0.7853	±	0.0323
- moral_disputes	Yaml	none	acc	0.7168	±	0.0243
- moral_scenarios	Yaml	none	acc	0.3207	±	0.0156
- philosophy	Yaml	none	acc	0.7042	±	0.0259
- prehistory	Yaml	none	acc	0.7593	±	0.0238
- professional_law	Yaml	none	acc	0.4433	±	0.0127
- world_religions	Yaml	none	acc	0.8363	±	0.0284
- other	N/A	none	acc	0.6987	±	0.1048
- business_ethics	Yaml	none	acc	0.5800	±	0.0496
- clinical_knowledge	Yaml	none	acc	0.7019	±	0.0282
- college_medicine	Yaml	none	acc	0.6474	±	0.0364
- global_facts	Yaml	none	acc	0.3900	±	0.0490
- human_aging	Yaml	none	acc	0.6502	±	0.0320
- management	Yaml	none	acc	0.7864	±	0.0406
- marketing	Yaml	none	acc	0.8590	±	0.0228
- medical_genetics	Yaml	none	acc	0.7400	±	0.0441
- miscellaneous	Yaml	none	acc	0.8148	±	0.0139
- nutrition	Yaml	none	acc	0.7418	±	0.0251
- professional_accounting	Yaml	none	acc	0.4929	±	0.0298
- professional_medicine	Yaml	none	acc	0.6618	±	0.0287
- virology	Yaml	none	acc	0.5482	±	0.0387
- social_sciences	N/A	none	acc	0.7361	±	0.0640
- econometrics	Yaml	none	acc	0.5000	±	0.0470
- high_school_geography	Yaml	none	acc	0.7727	±	0.0299
- high_school_government_and_politics	Yaml	none	acc	0.8808	±	0.0234
- high_school_macroeconomics	Yaml	none	acc	0.6667	±	0.0239
- high_school_microeconomics	Yaml	none	acc	0.6597	±	0.0308
- high_school_psychology	Yaml	none	acc	0.8202	±	0.0165
- human_sexuality	Yaml	none	acc	0.7939	±	0.0355
- professional_psychology	Yaml	none	acc	0.6716	±	0.0190
- public_relations	Yaml	none	acc	0.6636	±	0.0453
- security_studies	Yaml	none	acc	0.7551	±	0.0275
- sociology	Yaml	none	acc	0.8209	±	0.0271
- us_foreign_policy	Yaml	none	acc	0.8300	±	0.0378
- stem	N/A	none	acc	0.5268	±	0.1263
- abstract_algebra	Yaml	none	acc	0.3200	±	0.0469
- anatomy	Yaml	none	acc	0.6296	±	0.0417
- astronomy	Yaml	none	acc	0.6645	±	0.0384
- college_biology	Yaml	none	acc	0.7431	±	0.0365
- college_chemistry	Yaml	none	acc	0.4800	±	0.0502
- college_computer_science	Yaml	none	acc	0.5200	±	0.0502
- college_mathematics	Yaml	none	acc	0.4200	±	0.0496
- college_physics	Yaml	none	acc	0.4510	±	0.0495
- computer_security	Yaml	none	acc	0.7800	±	0.0416
- conceptual_physics	Yaml	none	acc	0.5489	±	0.0325
- electrical_engineering	Yaml	none	acc	0.5655	±	0.0413
- elementary_mathematics	Yaml	none	acc	0.3915	±	0.0251
- high_school_biology	Yaml	none	acc	0.7548	±	0.0245
- high_school_chemistry	Yaml	none	acc	0.5222	±	0.0351
- high_school_computer_science	Yaml	none	acc	0.6900	±	0.0465
- high_school_mathematics	Yaml	none	acc	0.3222	±	0.0285
- high_school_physics	Yaml	none	acc	0.3444	±	0.0388
- high_school_statistics	Yaml	none	acc	0.5139	±	0.0341
- machine_learning	Yaml	none	acc	0.4643	±	0.0473

Groups	Version	Filter	Metric	Value		Stderr
mmlu	N/A	none	acc	0.6236	±	0.1269
- humanities	N/A	none	acc	0.5651	±	0.1289
- other	N/A	none	acc	0.6987	±	0.1048
- social_sciences	N/A	none	acc	0.7361	±	0.0640
- stem	N/A	none	acc	0.5268	±	0.1263

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}, }