metadata

datasets:
  - HuggingFaceH4/ultrachat_200k
  - HuggingFaceH4/ultrafeedback_binarized
  - meta-math/MetaMathQA
  - WizardLM/WizardLM_evol_instruct_V2_196k
  - Intel/orca_dpo_pairs
language:
  - en
tags:
  - causal-lm
extra_gated_fields:
  Name: text
  Email: text
  Country: text
  Organization or Affiliation: text
  I ALLOW Stability AI to email me about new model releases: checkbox
license: other

`StableLM Zephyr 3B`

Model Description

StableLM Zephyr 3B is a 3 billion parameter instruction tuned inspired by HugginFaceH4's Zephyr 7B training pipeline this model was trained on a mix of publicly available datasets, synthetic datasets using Direct Preference Optimization (DPO), evaluation for this model based on MT Bench and Alpaca Benchmark

Usage

StableLM Zephyr 3B uses the following instruction format:

<|user|>
List 3 synonyms for the word "tiny"<|endoftext|>
<|assistant|>
1. Dwarf
2. Little
3. Petite<|endoftext|>

This format is also available through the tokenizer's apply_chat_template method:

from transformers import AutoModelForCausalLM, AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained('stabilityai/stablelm-zephyr-3b')
model = AutoModelForCausalLM.from_pretrained(
    'stabilityai/stablelm-zephyr-3b',
    trust_remote_code=True,
    device_map="auto"
)

prompt = [{'role': 'user', 'content': 'List 3 synonyms for the word "tiny"'}]
inputs = tokenizer.apply_chat_template(
    prompt,
    add_generation_prompt=True,
    return_tensors='pt'
)

tokens = model.generate(
    inputs.to(model.device),
    max_new_tokens=1024,
    temperature=0.8,
    do_sample=True
)

print(tokenizer.decode(tokens[0], skip_special_tokens=False))

Model Details

Developed by: Stability AI
Model type: StableLM Zephyr 3B model is an auto-regressive language model based on the transformer decoder architecture.
Language(s): English
Library: Alignment Handbook
Finetuned from model: stabilityai/stablelm-3b-4e1t
License: StabilityAI Non-Commercial Research Community License
Contact: For questions and comments about the model, please email [email protected]

Training Dataset

The dataset is comprised of a mixture of open datasets large-scale datasets available on the HuggingFace Hub:

SFT Datasets

HuggingFaceH4/ultrachat_200k
meta-math/MetaMathQA
WizardLM/WizardLM_evol_instruct_V2_196k
Open-Orca/SlimOrca

Preference Datasets:

HuggingFaceH4/ultrafeedback_binarized
Intel/orca_dpo_pairs

Performance

MT-Bench and Alpaca Bench

Model	Size	Alignment	MT-Bench (score)	AlpacaEval (win rate %)
StableLM Zephyr 3B 🪁	3B	DPO	6.64	76.00
StableLM Zephyr (SFT only)	3B	SFT	6.04	71.15
Capybara v1.9	3B	dSFT	5.94	-
MPT-Chat	7B	dSFT	5.42	-
Xwin-LM v0.1	7B	dPPO	6.19	87.83
Mistral-Instruct v0.1	7B	-	6.84	-
Zephyr-7b-α	7B	dDPO	6.88	-
Zephyr-7b-β	7B	dDPO	7.34	90.60
Falcon-Instruct	40B	dSFT	5.17	45.71
Guanaco	65B	SFT	6.41	71.80
Llama2-Chat	70B	RLHF	6.86	92.66
Vicuna v1.3	33B	dSFT	7.12	88.99
WizardLM v1.0	70B	dSFT	7.71	-
Xwin-LM v0.1	70B	dPPO	-	95.57
GPT-3.5-turbo	-	RLHF	7.94	89.37
Claude 2	-	RLHF	8.06	91.36
GPT-4	-	RLHF	8.99	95.28

Other benchmark:

HuggingFace OpenLLM Leaderboard

Metric Value

ARC (25-shot) 47.0

HellaSwag (10-shot) 74.2

MMLU (5-shot) 46.3

TruthfulQA (0-shot) 46.5

Winogrande (5-shot) 65.5

GSM8K (5-shot) 42.3
BigBench:

Metric	Value
ARC (25-shot)	47.0
HellaSwag (10-shot)	74.2
MMLU (5-shot)	46.3
TruthfulQA (0-shot)	46.5
Winogrande (5-shot)	65.5
GSM8K (5-shot)	42.3

Average: 35.26
Details:

Task	Metric	Value	Stderr
bigbench_causal_judgement	multiple_choice_grade	0.5316	0.0363
bigbench_date_understanding	multiple_choice_grade	0.4363	0.0259
bigbench_disambiguation_qa	multiple_choice_grade	0.3217	0.0291
bigbench_dyck_languages	multiple_choice_grade	0.1450	0.0111
bigbench_formal_fallacies_syllogisms_negation	multiple_choice_grade	0.4982	0.0042
bigbench_geometric_shapes	multiple_choice_grade	0.1086	0.0164
bigbench_hyperbaton	exact_str_match	0.0000	0.0000
bigbench_logical_deduction_five_objects	multiple_choice_grade	0.5232	0.0022
bigbench_logical_deduction_seven_objects	multiple_choice_grade	0.2480	0.0193
bigbench_logical_deduction_three_objects	multiple_choice_grade	0.1814	0.0146
bigbench_movie_recommendation	multiple_choice_grade	0.4067	0.0284
bigbench_navigate	multiple_choice_grade	0.2580	0.0196
bigbench_reasoning_about_colored_objects	multiple_choice_grade	0.5990	0.0155
bigbench_ruin_names	multiple_choice_grade	0.4370	0.0111
bigbench_salient_translation_error_detection	multiple_choice_grade	0.3951	0.0231
bigbench_snarks	multiple_choice_grade	0.2265	0.0133
bigbench_sports_understanding	multiple_choice_grade	0.6464	0.0356
bigbench_temporal_sequences	multiple_choice_grade	0.5091	0.0159
bigbench_tracking_shuffled_objects_five_objects	multiple_choice_grade	0.2680	0.0140
bigbench_tracking_shuffled_objects_seven_objects	multiple_choice_grade	0.1856	0.0110
bigbench_tracking_shuffled_objects_three_objects	multiple_choice_grade	0.1269	0.0080

AGI Benchmark:

Average: 33.23
Details: | Task |Version| Metric |Value | |Stderr|

|------------------------------|------:|--------|-----:|---|-----:| |agieval_aqua_rat | 0|acc |0.2126|± |0.0257| | | |acc_norm|0.1890|± |0.0246| |agieval_gaokao_biology | 0|acc |0.2571|± |0.0302| | | |acc_norm|0.3143|± |0.0321| |agieval_gaokao_chemistry | 0|acc |0.2464|± |0.0300| | | |acc_norm|0.2899|± |0.0316| |agieval_gaokao_chinese | 0|acc |0.2927|± |0.0291| | | |acc_norm|0.3049|± |0.0294| |agieval_gaokao_english | 0|acc |0.6176|± |0.0278| | | |acc_norm|0.6438|± |0.0274| |agieval_gaokao_geography | 0|acc |0.3015|± |0.0326| | | |acc_norm|0.3065|± |0.0328| |agieval_gaokao_history | 0|acc |0.3106|± |0.0303| | | |acc_norm|0.3319|± |0.0308| |agieval_gaokao_mathqa | 0|acc |0.2650|± |0.0236| | | |acc_norm|0.2707|± |0.0237| |agieval_gaokao_physics | 0|acc |0.3450|± |0.0337| | | |acc_norm|0.3550|± |0.0339| |agieval_logiqa_en | 0|acc |0.2980|± |0.0179| | | |acc_norm|0.3195|± |0.0183| |agieval_logiqa_zh | 0|acc |0.2842|± |0.0177| | | |acc_norm|0.3318|± |0.0185| |agieval_lsat_ar | 0|acc |0.2000|± |0.0264| | | |acc_norm|0.2043|± |0.0266| |agieval_lsat_lr | 0|acc |0.3176|± |0.0206| | | |acc_norm|0.3275|± |0.0208| |agieval_lsat_rc | 0|acc |0.4312|± |0.0303| | | |acc_norm|0.4201|± |0.0301| |agieval_sat_en | 0|acc |0.6117|± |0.0340| | | |acc_norm|0.6117|± |0.0340| |agieval_sat_en_without_passage| 0|acc |0.3398|± |0.0331| | | |acc_norm|0.3495|± |0.0333| |agieval_sat_math | 0|acc |0.3182|± |0.0315| | | |acc_norm|0.2909|± |0.0307|

Training Infrastructure

Hardware: StableLM Zephyr 3B was trained on the Stability AI cluster across 8 nodes with 8 A100 80GBs GPUs for each nodes.
Code Base: We use our internal script for SFT steps and used HuggingFace Alignment Handbook script for DPO training.

Commitment to Ethical AI

In line with our responsibility towards ethical AI development, StableLM Zephyr 3B is released with a focus on ensuring safety, reliability, and appropriateness in its applications. To this end, we have evaluated StableLM Zephyr 3B on 488 malicious prompts and used standard protocols to assess the harmfulness of its outputs. Compared to Zephyr-7b-β, StableLM Zephyr 3B reduces the number of harmful outputs as assessed by GPT-4 by 55. Additionally, we performed an internal red teaming event targeting the following abuse areas:

Self-Harm Methods: (Suicide Methods, Encouragement of Self-Harm, Methods and encouragement of Eating Disorders)
Misinformation: (Health, Conspiracy Theories, Social Unrest/Conflict, Political Misinformation, & Climate change)
Hate Speech: (Race, Stereotypes, Immigrants, Gender, Personally Identifiable Information such as Social security numbers, Full names, ID numbers, Email addresses, and telephone numbers)

We have incorporated the findings of our malicious prompts evaluation and red teaming event into our release. Users are encouraged to fine-tune and evaluate the model to suit their specific needs, considering the potential biases and limitations found in StableLM Zephyr 3B and inherent in other LLM models.

Use and Limitations

Intended Use

The model is intended to be used as a foundational base model for application-specific fine-tuning. Developers must evaluate and fine-tune the model for safe performance in downstream applications.

Limitations and Bias

This model is not trained against adversarial inputs. We strongly recommend pairing this model with an input and output classifier to prevent harmful responses.

Through our internal red teaming, we discovered that while the model will not output harmful information if not prompted to do so, it is willing to output potentially harmful outputs or misinformation when the user requests it. Using this model will require guardrails around your inputs and outputs to ensure that any outputs returned are not misinformation or harmful. Additionally, as each use case is unique, we recommend running your own suite of tests to ensure proper performance of this model. Finally, do not use the models if they are unsuitable for your application, or for any applications that may cause deliberate or unintentional harm to others.