Phi-3.5-mini is a lightweight, state-of-the-art open model built upon datasets used for Phi-3 - synthetic data and filtered publicly available websites - with a focus on very high-quality, reasoning dense data. The model belongs to the Phi-3 model family and supports 128K token context length. The model underwent a rigorous enhancement process, incorporating both 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.
Release Notes
This is an update over the June 2024 instruction-tuned Phi-3 Mini release based on valuable user feedback. The model used additional post-training data leading to substantial gains on multilingual, multi-turn conversation quality, and reasoning capability. We believe most use cases will benefit from this release, but we encourage users to test in their particular AI applications. We appreciate the enthusiastic adoption of the Phi-3 model family, and continue to welcome all feedback from the community.
Multilingual
The table below highlights multilingual capability of the Phi-3.5 Mini on multilingual MMLU, MEGA, and multilingual MMLU-pro datasets. Overall, we observed that even with just 3.8B active parameters, the model is competitive on multilingual tasks in comparison to other models with a much bigger active parameters.
Benchmark
Phi-3.5 Mini-Ins
Phi-3.1-Mini-128K-Ins
Mistral-7B-Instruct-v0.3
Mistral-Nemo-12B-Ins-2407
Llama-3.1-8B-Ins
Gemma-2-9B-Ins
Gemini 1.5 Flash
GPT-4o-mini-2024-07-18 (Chat)
Multilingual MMLU
55.4
51.08
47.4
58.9
56.2
63.8
77.2
72.9
Multilingual MMLU-Pro
30.9
30.21
15.0
34.0
21.4
43.0
57.9
53.2
MGSM
47.9
41.56
31.8
63.3
56.7
75.1
75.8
81.7
MEGA MLQA
61.7
55.5
43.9
61.2
45.2
54.4
61.6
70.0
MEGA TyDi QA
62.2
55.9
54.0
63.7
54.5
65.6
63.6
81.8
MEGA UDPOS
46.5
48.1
57.2
58.2
54.1
56.6
62.4
66.0
MEGA XCOPA
63.1
62.4
58.8
10.8
21.1
31.2
95.0
90.3
MEGA XStoryCloze
73.5
73.6
75.5
92.3
71.0
87.0
20.7
96.6
Average
55.2
52.3
47.9
55.3
47.5
59.6
64.3
76.6
The table below shows Multilingual MMLU scores in some of the supported languages. For more multi-lingual benchmarks and details, see Appendix A.
Benchmark
Phi-3.5 Mini-Ins
Phi-3.1-Mini-128K-Ins
Mistral-7B-Instruct-v0.3
Mistral-Nemo-12B-Ins-2407
Llama-3.1-8B-Ins
Gemma-2-9B-Ins
Gemini 1.5 Flash
GPT-4o-mini-2024-07-18 (Chat)
Arabic
44.2
35.4
33.7
45.3
49.1
56.3
73.6
67.1
Chinese
52.6
46.9
45.9
58.2
54.4
62.7
66.7
70.8
Dutch
57.7
48.0
51.3
60.1
55.9
66.7
80.6
74.2
French
61.1
61.7
53.0
63.8
62.8
67.0
82.9
75.6
German
62.4
61.3
50.1
64.5
59.9
65.7
79.5
74.3
Italian
62.8
63.1
52.5
64.1
55.9
65.7
82.6
75.9
Russian
50.4
45.3
48.9
59.0
57.4
63.2
78.7
72.6
Spanish
62.6
61.3
53.9
64.3
62.6
66.0
80.0
75.5
Ukrainian
45.2
36.7
46.9
56.6
52.9
62.0
77.4
72.6
Long Context
Phi-3.5-mini supports 128K context length, therefore the model is capable of several long context tasks including long document/meeting summarization, long document QA, long document information retrieval. We see that Phi-3.5-mini is clearly better than Gemma-2 family which only supports 8K context length. Phi-3.5-mini is competitive with other much larger open-weight models such as Llama-3.1-8B-instruct, Mistral-7B-instruct-v0.3, and Mistral-Nemo-12B-instruct-2407.
Benchmark
Phi-3.5-mini-instruct
Llama-3.1-8B-instruct
Mistral-7B-instruct-v0.3
Mistral-Nemo-12B-instruct-2407
Gemini-1.5-Flash
GPT-4o-mini-2024-07-18 (Chat)
GovReport
25.9
25.1
26.0
25.6
27.8
24.8
QMSum
21.3
21.6
21.3
22.1
24.0
21.7
Qasper
41.9
37.2
31.4
30.7
43.5
39.8
SQuALITY
25.3
26.2
25.9
25.8
23.5
23.8
SummScreenFD
16.0
17.6
17.5
18.2
16.3
17.0
Average
26.1
25.5
24.4
24.5
27.0
25.4
RULER: a retrieval-based benchmark for long context understanding
Model
4K
8K
16K
32K
64K
128K
Average
Phi-3.5-mini-instruct
94.3
91.1
90.7
87.1
78.0
63.6
84.1
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-mini-instruct
86
67
73
77
82
77
Llama-3.1-8B-instruct
80
65
73
76
63
71
Mistral-7B-instruct-v0.3
61
57
51
61
80
62
Usage
Requirements
Phi-3 family has been integrated in the 4.43.0 version of transformers. The current transformers version can be verified with: pip list | grep transformers.
Phi-3.5-mini-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-mini-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-mini-instruct model checkpoint, 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-mini-instruct",8 device_map="cuda",9 torch_dtype="auto",10 trust_remote_code=True,11)12tokenizer = AutoTokenizer.from_pretrained("microsoft/Phi-3.5-mini-instruct")1314messages =[15{"role":"system","content":"You are a helpful AI assistant."},16{"role":"user","content":"Can you provide ways to eat combinations of bananas and dragonfruits?"},17{"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."},18{"role":"user","content":"What about solving an 2x + 3 = 7 equation?"},19]2021pipe = pipeline(22"text-generation",23 model=model,24 tokenizer=tokenizer,25)2627generation_args ={28"max_new_tokens":500,29"return_full_text":False,30"temperature":0.0,31"do_sample":False,32}3334output = pipe(messages,**generation_args)35print(output[0]['generated_text'])
Notes: If you want to use flash attention, call AutoModelForCausalLM.from_pretrained() with attn_implementation="flash_attention_2"
Responsible AI Considerations
Like other language models, the Phi series 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. English language varieties with less representation in the training data might experience worse performance than standard American English.
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 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.
Developers should apply responsible AI best practices and are responsible for ensuring that a specific use case complies with relevant laws and regulations (e.g. privacy, trade, etc.). 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.
Training
Model
Architecture: Phi-3.5-mini has 3.8B parameters and is a dense decoder-only Transformer model using the same tokenizer as Phi-3 Mini. Inputs: Text. It is best suited for prompts using chat format. Context length: 128K tokens GPUs: 512 H100-80G Training time: 10 days Training data: 3.4T tokens Outputs: Generated text in response to the input Dates: Trained between June 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 3.4 trillion tokens, 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.
Fine-tuning
A basic example of multi-GPUs supervised fine-tuning (SFT) with TRL and Accelerate modules is provided here.
Benchmarks
We report the results under completion format for Phi-3.5-mini on standard open-source benchmarks measuring the model's reasoning ability (both common sense reasoning and logical reasoning). We compare to Mistral-7B-Instruct-v0.3, Mistral-Nemo-12B-Ins-2407, Llama-3.1-8B-Ins, Gemma-2-9B-Ins, Gemini 1.5 Flash, and GPT-4o-mini-2024-07-18 (Chat).
All the reported numbers are produced with the exact same pipeline to ensure that the numbers are comparable. These numbers might differ from other published numbers due to slightly different choices in the evaluation.
As is now standard, we use few-shot prompts to evaluate the models, at temperature 0.
The prompts and number of shots are part of a Microsoft internal tool to evaluate language models, and in particular we did no optimization to the pipeline for Phi-3.
More specifically, we do not change prompts, pick different few-shot examples, change prompt format, or do any other form of optimization for the model.
The number of k–shot examples is listed per-benchmark. At the high-level overview of the model quality on representative benchmarks:
Category
Benchmark
Phi-3.5 Mini-Ins
Mistral-7B-Instruct-v0.3
Mistral-Nemo-12B-Ins-2407
Llama-3.1-8B-Ins
Gemma-2-9B-Ins
Gemini 1.5 Flash
GPT-4o-mini-2024-07-18 (Chat)
Popular aggregated benchmark
Arena Hard
37
18.1
39.4
25.7
42
55.2
75
BigBench Hard CoT (0-shot)
69
33.4
60.2
63.4
63.5
66.7
80.4
MMLU (5-shot)
69
60.3
67.2
68.1
71.3
78.7
77.2
MMLU-Pro (0-shot, CoT)
47.4
18
40.7
44
50.1
57.2
62.8
Reasoning
ARC Challenge (10-shot)
84.6
77.9
84.8
83.1
89.8
92.8
93.5
BoolQ (2-shot)
78
80.5
82.5
82.8
85.7
85.8
88.7
GPQA (0-shot, CoT)
30.4
15.6
28.6
26.3
29.2
37.5
41.1
HellaSwag (5-shot)
69.4
71.6
76.7
73.5
80.9
67.5
87.1
OpenBookQA (10-shot)
79.2
78
84.4
84.8
89.6
89
90
PIQA (5-shot)
81
73.4
83.5
81.2
83.7
87.5
88.7
Social IQA (5-shot)
74.7
73
75.3
71.8
74.7
77.8
82.9
TruthfulQA (MC2) (10-shot)
64
64.7
68.1
69.2
76.6
76.6
78.2
WinoGrande (5-shot)
68.5
58.1
70.4
64.7
74
74.7
76.9
Multilingual
Multilingual MMLU (5-shot)
55.4
47.4
58.9
56.2
63.8
77.2
72.9
MGSM (0-shot CoT)
47.9
31.8
63.3
56.7
76.4
75.8
81.7
Math
GSM8K (8-shot, CoT)
86.2
54.4
84.2
82.4
84.9
82.4
91.3
MATH (0-shot, CoT)
48.5
19
31.2
47.6
50.9
38
70.2
Long context
Qasper
41.9
31.4
30.7
37.2
13.9
43.5
39.8
SQuALITY
24.3
25.9
25.8
26.2
0
23.5
23.8
Code Generation
HumanEval (0-shot)
62.8
35.4
63.4
66.5
61
74.4
86.6
MBPP (3-shot)
69.6
50.4
68.1
69.4
69.3
77.5
84.1
Average
61.4
48.5
61.3
61.0
63.3
68.5
74.9
We take a closer look at different categories across public benchmark datasets at the table below:
Category
Phi-3.5 Mini-Ins
Mistral-7B-Instruct-v0.3
Mistral-Nemo-12B-Ins-2407
Llama-3.1-8B-Ins
Gemma-2-9B-Ins
Gemini 1.5 Flash
GPT-4o-mini-2024-07-18 (Chat)
Popular aggregated benchmark
55.6
32.5
51.9
50.3
56.7
64.5
73.9
Reasoning
70.1
65.2
72.2
70.5
75.4
77.7
80
Language understanding
62.6
62.8
67
62.9
72.8
66.6
76.8
Robustness
59.7
53.4
65.2
59.8
64.7
68.9
77.5
Long context
26.1
25.5
24.4
24.5
0
27
25.4
Math
67.4
36.7
57.7
65
67.9
60.2
80.8
Code generation
62
43.1
56.9
65.8
58.3
66.8
69.9
Multilingual
55.2
47.9
55.3
47.5
59.6
64.3
76.6
Overall, the model with only 3.8B-param achieves a similar level of multilingual language understanding and reasoning ability as much larger models. 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.
Note that by default, the Phi-3.5-mini-instruct model uses flash attention, which requires certain types of GPU hardware to run. We have tested on the following GPU types:
NVIDIA A100
NVIDIA A6000
NVIDIA H100
If you want to run the model on:
NVIDIA V100 or earlier generation GPUs: call AutoModelForCausalLM.from_pretrained() with attn_implementation="eager"
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.
