Phi-3.5-MoE is a lightweight, state-of-the-art open model built upon datasets used for Phi-3 - synthetic data and filtered publicly available documents - with a focus on very high-quality, reasoning dense data. The model supports multilingual and comes with 128K context length (in tokens). The model underwent a rigorous enhancement process, incorporating supervised fine-tuning, proximal policy optimization, and direct preference optimization to ensure precise instruction adherence and robust safety measures.
The model is intended for commercial and research use in multiple languages. The model provides uses for general purpose AI systems and applications which require:
Memory/compute constrained environments
Latency bound scenarios
Strong reasoning (especially code, math and logic)
Our model is designed to accelerate research on language and multimodal models, for use as a building block for generative AI powered features.
Use Case Considerations
Our models are not specifically designed or evaluated for all downstream purposes. Developers should consider common limitations of language models as they select use cases, and evaluate and mitigate for accuracy, safety, and fariness before using within a specific downstream use case, particularly for high risk scenarios. Developers should be aware of and adhere to applicable laws or regulations (including privacy, trade compliance laws, etc.) that are relevant to their use case.
Nothing contained in this Model Card should be interpreted as or deemed a restriction or modification to the license the model is released under.
Usage
Requirements
Phi-3.5-MoE-instruct is integrated in the official version of transformers starting from 4.46.0.
The current transformers version can be verified with: pip list | grep transformers.
Phi-3.5-MoE-Instruct supports a vocabulary size of up to 32064 tokens. The tokenizer files already provide placeholder tokens that can be used for downstream fine-tuning, but they can also be extended up to the model's vocabulary size.
Input Formats
Given the nature of the training data, the Phi-3.5-MoE-instruct model is best suited for prompts using the chat format as follows:
<|system|>
You are a helpful assistant.<|end|>
<|user|>
How to explain Internet for a medieval knight?<|end|>
<|assistant|>
Loading the model locally
After obtaining the Phi-3.5-MoE-instruct model checkpoints, users can use this sample code for inference.
python
1import torch
2from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline
34torch.random.manual_seed(0)56model = AutoModelForCausalLM.from_pretrained(7"microsoft/Phi-3.5-MoE-instruct",8 device_map="cuda",9 torch_dtype="auto",10 trust_remote_code=False,11)1213tokenizer = AutoTokenizer.from_pretrained("microsoft/Phi-3.5-MoE-instruct")1415messages =[16{"role":"system","content":"You are a helpful AI assistant."},17{"role":"user","content":"Can you provide ways to eat combinations of bananas and dragonfruits?"},18{"role":"assistant","content":"Sure! Here are some ways to eat bananas and dragonfruits together: 1. Banana and dragonfruit smoothie: Blend bananas and dragonfruits together with some milk and honey. 2. Banana and dragonfruit salad: Mix sliced bananas and dragonfruits together with some lemon juice and honey."},19{"role":"user","content":"What about solving an 2x + 3 = 7 equation?"},20]2122pipe = pipeline(23"text-generation",24 model=model,25 tokenizer=tokenizer,26)2728generation_args ={29"max_new_tokens":500,30"return_full_text":False,31"temperature":0.0,32"do_sample":False,33}3435output = pipe(messages,**generation_args)36print(output[0]['generated_text'])
Benchmarks
To understand the capabilities, we compare Phi-3.5-MoE with a set of models over a variety of benchmarks using our internal benchmark platform. At the high-level overview of the model quality on representative benchmarks:
Category
Benchmark
Phi-3.5-MoE-instruct
Mistral-Nemo-12B-instruct-2407
Llama-3.1-8B-instruct
Gemma-2-9b-It
Gemini-1.5-Flash
GPT-4o-mini-2024-07-18 (Chat)
Popular aggregated benchmark
Arena Hard
37.9
39.4
25.7
42.0
55.2
75.0
BigBench Hard CoT (0-shot)
79.1
60.2
63.4
63.5
66.7
80.4
MMLU (5-shot)
78.9
67.2
68.1
71.3
78.7
77.2
MMLU-Pro (0-shot, CoT)
54.3
40.7
44.0
50.1
57.2
62.8
Reasoning
ARC Challenge (10-shot)
91.0
84.8
83.1
89.8
92.8
93.5
BoolQ (2-shot)
84.6
82.5
82.8
85.7
85.8
88.7
GPQA (0-shot, CoT)
36.8
28.6
26.3
29.2
37.5
41.1
HellaSwag (5-shot)
83.8
76.7
73.5
80.9
67.5
87.1
OpenBookQA (10-shot)
89.6
84.4
84.8
89.6
89.0
90.0
PIQA (5-shot)
88.6
83.5
81.2
83.7
87.5
88.7
Social IQA (5-shot)
78.0
75.3
71.8
74.7
77.8
82.9
TruthfulQA (MC2) (10-shot)
77.5
68.1
69.2
76.6
76.6
78.2
WinoGrande (5-shot)
81.3
70.4
64.7
74.0
74.7
76.9
Multilingual
Multilingual MMLU (5-shot)
69.9
58.9
56.2
63.8
77.2
72.9
MGSM (0-shot CoT)
58.7
63.3
56.7
75.1
75.8
81.7
Math
GSM8K (8-shot, CoT)
88.7
84.2
82.4
84.9
82.4
91.3
MATH (0-shot, CoT)
59.5
31.2
47.6
50.9
38.0
70.2
Long context
Qasper
40.0
30.7
37.2
13.9
43.5
39.8
SQuALITY
24.1
25.8
26.2
0.0
23.5
23.8
Code Generation
HumanEval (0-shot)
70.7
63.4
66.5
61.0
74.4
86.6
MBPP (3-shot)
80.8
68.1
69.4
69.3
77.5
84.1
Average
69.2
61.3
61.0
63.3
68.5
74.9
We take a closer look at different categories across 80 public benchmark datasets at the table below:
Category
Phi-3.5-MoE-instruct
Mistral-Nemo-12B-instruct-2407
Llama-3.1-8B-instruct
Gemma-2-9b-It
Gemini-1.5-Flash
GPT-4o-mini-2024-07-18 (Chat)
Popular aggregated benchmark
62.6
51.9
50.3
56.7
64.5
73.9
Reasoning
78.7
72.2
70.5
75.4
77.7
80.0
Language understanding
71.8
67.0
62.9
72.8
66.6
76.8
Robustness
75.6
65.2
59.8
64.7
68.9
77.5
Long context
25.5
24.5
25.5
0.0
27.0
25.4
Math
74.1
57.7
65.0
67.9
60.2
80.8
Code generation
68.3
56.9
65.8
58.3
66.8
69.9
Multilingual
65.8
55.3
47.5
59.6
64.3
76.6
Overall, Phi-3.5-MoE with only 6.6B active parameters achieves a similar level of language understanding and math as much larger models. Moreover, the model outperforms bigger models in reasoning capability and only behind GPT-4o-mini. However, it is still fundamentally limited by its size for certain tasks. The model simply does not have the capacity to store too much factual knowledge, therefore, users may experience factual incorrectness. However, we believe such weakness can be resolved by augmenting Phi-3.5 with a search engine, particularly when using the model under RAG settings.
Multilingual
The table below highlights multilingual capability of Phi-3.5-MoE on multilingual MMLU, MEGA, and multilingual MMLU-pro datasets. Overall, we observed that even with just 6.6B active parameters, the model is very competitive on multilingual tasks in comparison to other models with a much bigger active parameters.
Category
Phi-3.5-MoE-instruct
Mistral-Nemo-12B-instruct-2407
Llama-3.1-8B-instruct
Gemma-2-9b-It
Gemini-1.5-Flash
GPT-4o-mini-2024-07-18 (Chat)
Multilingual MMLU
69.9
58.9
56.2
63.8
77.2
72.9
Multilingual MMLU-Pro
45.3
34.0
21.4
43.0
57.9
53.2
MGSM
58.7
63.3
56.7
75.1
75.8
81.7
MEGA MLQA
65.3
61.2
45.2
54.4
61.6
70.0
MEGA TyDi QA
67.1
63.7
54.5
65.6
63.6
81.8
MEGA UDPOS
60.4
58.2
54.1
56.6
62.4
66.0
MEGA XCOPA
76.6
10.8
21.1
31.2
95.0
90.3
MEGA XStoryCloze
82.8
92.3
71.0
87.0
20.7
96.6
Average
65.8
55.3
47.5
59.6
64.3
76.6
Long Context
Phi-3.5-MoE supports 128K context length, therefore the model is capable of several long context tasks including long document/meeting summarization, long document QA, multilingual context retrieval. We see that Phi-3.5 is clearly better than Gemma-2 family which only supports 8K context length. Phi-3.5-MoE-instruct is very competitive with other much larger open-weight models such as Llama-3.1-8B-instruct, and Mistral-Nemo-12B-instruct-2407.
Benchmark
Phi-3.5-MoE-instruct
Mistral-Nemo-12B-instruct-2407
Llama-3.1-8B-instruct
Gemini-1.5-Flash
GPT-4o-mini-2024-07-18 (Chat)
GovReport
26.4
25.6
25.1
27.8
24.8
QMSum
19.9
22.1
21.6
24.0
21.7
Qasper
40.0
30.7
37.2
43.5
39.8
SQuALITY
24.1
25.8
26.2
23.5
23.8
SummScreenFD
16.9
18.2
17.6
16.3
17.0
Average
25.5
24.5
25.5
27.0
25.4
RULER: a retrieval-based benchmark for long context understanding
Model
4K
8K
16K
32K
64K
128K
Average
Phi-3.5-MoE-instruct
94.8
93
93.2
91.6
85.7
64.2
87.1
Llama-3.1-8B-instruct
95.5
93.8
91.6
87.4
84.7
77.0
88.3
Mistral-Nemo-12B-instruct-2407
87.8
87.2
87.7
69.0
46.8
19.0
66.2
RepoQA: a benchmark for long context code understanding
Model
Python
C++
Rust
Java
TypeScript
Average
Phi-3.5-MoE-instruct
89
74
81
88
95
85
Llama-3.1-8B-instruct
80
65
73
76
63
71
Mistral-7B-instruct-v0.3
61
57
51
61
80
62
Training
Model
Architecture: Phi-3.5-MoE has 16x3.8B parameters with 6.6B active parameters when using 2 experts. The model is a mixture-of-expert decoder-only Transformer model using the tokenizer with vocabulary size of 32,064. Inputs: Text. It is best suited for prompts using chat format. Context length: 128K tokens GPUs: 512 H100-80G Training time: 23 days Training data: 4.9T tokens Outputs: Generated text in response to the input Dates: Trained between April and August 2024 Status: This is a static model trained on an offline dataset with cutoff date October 2023 for publicly available data. Future versions of the tuned models may be released as we improve models. Supported languages: Arabic, Chinese, Czech, Danish, Dutch, English, Finnish, French, German, Hebrew, Hungarian, Italian, Japanese, Korean, Norwegian, Polish, Portuguese, Russian, Spanish, Swedish, Thai, Turkish, Ukrainian Release date: August 2024
Training Datasets
Our training data includes a wide variety of sources, totaling 4.9 trillion tokens (including 10% multilingual), and is a combination of
publicly available documents filtered rigorously for quality, selected high-quality educational data, and code;
newly created synthetic, “textbook-like” data for the purpose of teaching math, coding, common sense reasoning, general knowledge of the world (science, daily activities, theory of mind, etc.);
high quality chat format supervised data covering various topics to reflect human preferences on different aspects such as instruct-following, truthfulness, honesty and helpfulness.
We are focusing on the quality of data that could potentially improve the reasoning ability for the model, and we filter the publicly available documents to contain the correct level of knowledge. As an example, the result of a game in premier league in a particular day might be good training data for frontier models, but we need to remove such information to leave more model capacity for reasoning for the small size models. More details about data can be found in the Phi-3 Technical Report.
Responsible AI Considerations
Like other language models, the Phi family of models can potentially behave in ways that are unfair, unreliable, or offensive. Some of the limiting behaviors to be aware of include:
Quality of Service: The Phi models are trained primarily on English text and some additional multilingual text. Languages other than English will experience worse performance as well as performance disparities across non-English. English language varieties with less representation in the training data might experience worse performance than standard American English.
Multilingual performance and safety gaps: We believe it is important to make language models more widely available across different languages, but the Phi 3 models still exhibit challenges common across multilingual releases. As with any deployment of LLMs, developers will be better positioned to test for performance or safety gaps for their linguistic and cultural context and customize the model with additional fine-tuning and appropriate safeguards.
Representation of Harms & Perpetuation of Stereotypes: These models can over- or under-represent groups of people, erase representation of some groups, or reinforce demeaning or negative stereotypes. Despite safety post-training, these limitations may still be present due to differing levels of representation of different groups, cultural contexts, or prevalence of examples of negative stereotypes in training data that reflect real-world patterns and societal biases.
Inappropriate or Offensive Content: These models may produce other types of inappropriate or offensive content, which may make it inappropriate to deploy for sensitive contexts without additional mitigations that are specific to the use case.
Information Reliability: Language models can generate nonsensical content or fabricate content that might sound reasonable but is inaccurate or outdated.
Limited Scope for Code: Majority of Phi-3 training data is based in Python and use common packages such as "typing, math, random, collections, datetime, itertools". If the model generates Python scripts that utilize other packages or scripts in other languages, we strongly recommend users manually verify all API uses.
Long Conversation: Phi-3 models, like other models, can in some cases generate responses that are repetitive, unhelpful, or inconsistent in very long chat sessions in both English and non-English languages. Developers are encouraged to place appropriate mitigations, like limiting conversation turns to account for the possible conversational drift
Developers should apply responsible AI best practices, including mapping, measuring, and mitigating risks associated with their specific use case and cultural, linguistic context. Phi-3 family of models are general purpose models. As developers plan to deploy these models for specific use cases, they are encouraged to fine-tune the models for their use case and leverage the models as part of broader AI systems with language-specific safeguards in place. Important areas for consideration include:
Allocation: Models may not be suitable for scenarios that could have consequential impact on legal status or the allocation of resources or life opportunities (ex: housing, employment, credit, etc.) without further assessments and additional debiasing techniques.
High-Risk Scenarios: Developers should assess the suitability of using models in high-risk scenarios where unfair, unreliable or offensive outputs might be extremely costly or lead to harm. This includes providing advice in sensitive or expert domains where accuracy and reliability are critical (ex: legal or health advice). Additional safeguards should be implemented at the application level according to the deployment context.
Misinformation: Models may produce inaccurate information. Developers should follow transparency best practices and inform end-users they are interacting with an AI system. At the application level, developers can build feedback mechanisms and pipelines to ground responses in use-case specific, contextual information, a technique known as Retrieval Augmented Generation (RAG).
Generation of Harmful Content: Developers should assess outputs for their context and use available safety classifiers or custom solutions appropriate for their use case.
Misuse: Other forms of misuse such as fraud, spam, or malware production may be possible, and developers should ensure that their applications do not violate applicable laws and regulations.
Safety Evaluation and Red-Teaming
We leveraged various evaluation techniques including red teaming, adversarial conversation simulations, and multilingual safety evaluation benchmark datasets to
evaluate Phi-3.5 models' propensity to produce undesirable outputs across multiple languages and risk categories.
Several approaches were used to compensate for the limitations of one approach alone. Findings across the various evaluation methods indicate that safety
post-training that was done as detailed in the Phi-3 Safety Post-Training paper had a positive impact across multiple languages and risk categories as observed by
refusal rates (refusal to output undesirable outputs) and robustness to jailbreak techniques. Note, however, while comprehensive red team evaluations were conducted
across all models in the prior release of Phi models, red teaming was largely focused on Phi-3.5 MOE across multiple languages and risk categories for this release as
it is the largest and more capable model of the three models. Details on prior red team evaluations across Phi models can be found in the Phi-3 Safety Post-Training paper.
For this release, insights from red teaming indicate that the models may refuse to generate undesirable outputs in English, even when the request for undesirable output
is in another language. Models may also be more susceptible to longer multi-turn jailbreak techniques across both English and non-English languages. These findings
highlight the need for industry-wide investment in the development of high-quality safety evaluation datasets across multiple languages, including low resource languages,
and risk areas that account for cultural nuances where those languages are spoken.
Note that by default, the Phi-3.5-MoE-instruct model uses flash attention, which requires certain types of GPU hardware to run. We have tested on the following GPU types:
This project may contain trademarks or logos for projects, products, or services. Authorized use of Microsoft trademarks or logos is subject to and must follow Microsoft’s Trademark & Brand Guidelines. Use of Microsoft trademarks or logos in modified versions of this project must not cause confusion or imply Microsoft sponsorship. Any use of third-party trademarks or logos are subject to those third-party’s policies.
Appendix A: Korean benchmarks
The prompt is the same as the CLIcK paper prompt. The experimental results below were given with max_tokens=512 (zero-shot), max_tokens=1024 (5-shot), temperature=0.01. No system prompt used.
GPT-4o: 2024-05-13 version
GPT-4o-mini: 2024-07-18 version
GPT-4-turbo: 2024-04-09 version
GPT-3.5-turbo: 2023-06-13 version
Overall, the Phi-3.5 MoE model with just 6.6B active params outperforms GPT-3.5-Turbo.
Benchmarks
Phi-3.5-MoE-Instruct
Phi-3.0-Mini-128k-Instruct (June2024)
Llama-3.1-8B-Instruct
GPT-4o
GPT-4o-mini
GPT-4-turbo
GPT-3.5-turbo
CLIcK
56.44
29.12
47.82
80.46
68.5
72.82
50.98
HAERAE 1.0
61.83
36.41
53.9
85.7
76.4
77.76
52.67
KMMLU (0-shot, CoT)
47.43
30.82
38.54
64.26
52.63
58.75
40.3
KMMLU (5-shot)
47.92
29.98
20.21
64.28
51.62
59.29
42.28
KMMLU-HARD (0-shot, CoT)
25.34
25.68
24.03
39.62
24.56
30.56
20.97
KMMLU-HARD (5-shot)
25.66
25.73
15.81
40.94
24.63
31.12
21.19
Average
45.82
29.99
29.29
62.54
50.08
56.74
39.61
CLIcK (Cultural and Linguistic Intelligence in Korean)
