metadata

license: other
license_link: https://huggingface.co/microsoft/phi-4/resolve/main/LICENSE
language:
  - en
pipeline_tag: text-generation
tags:
  - phi
  - nlp
  - math
  - code
  - chat
  - conversational
inference:
  parameters:
    temperature: 0
widget:
  - messages:
      - role: user
        content: How should I explain the Internet?
library_name: transformers

THIS IS A MIRROR OF https://ai.azure.com/explore/models/Phi-4/ ALONG WITH A CONVERTED TOKENIZER FOR llama.cpp

... OK tokenizer seems a bit off

OK, tokenizer seems a bit off 😂 (llama.cpp) root at nas in /mnt/llm/models llama-cli -m phi-4.etf16-Q6_K.gguf -p "Tell me a joke." -n 256 -t 8 -c 2048 --temp 0.8 -ngl 99 ggml_cuda_init: GGML_CUDA_FORCE_MMQ: no ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no ggml_cuda_init: found 2 CUDA devices: Device 0: NVIDIA GeForce RTX 3090, compute capability 8.6, VMM: yes Device 1: NVIDIA GeForce RTX 3090, compute capability 8.6, VMM: yes build: 1153 (d583cd03) with cc (GCC) 14.2.1 20240912 (Red Hat 14.2.1-3) for x86_64-redhat-linux main: llama backend init main: load the model and apply lora adapter, if any llama_load_model_from_file: using device CUDA0 (NVIDIA GeForce RTX 3090) - 24111 MiB free llama_load_model_from_file: using device CUDA1 (NVIDIA GeForce RTX 3090) - 24111 MiB free llama_model_loader: loaded meta data with 29 key-value pairs and 243 tensors from phi-4.etf16-Q6_K.gguf (version GGUF V3 (latest)) llama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output. llama_model_loader: - kv 0: general.architecture str = phi3 llama_model_loader: - kv 1: general.type str = model llama_model_loader: - kv 2: general.name str = Phi 4 llama_model_loader: - kv 3: general.version str = 4 llama_model_loader: - kv 4: general.organization str = Microsoft llama_model_loader: - kv 5: general.basename str = phi llama_model_loader: - kv 6: general.size_label str = 15B llama_model_loader: - kv 7: phi3.context_length u32 = 16384 llama_model_loader: - kv 8: phi3.rope.scaling.original_context_length u32 = 16384 llama_model_loader: - kv 9: phi3.embedding_length u32 = 5120 llama_model_loader: - kv 10: phi3.feed_forward_length u32 = 17920 llama_model_loader: - kv 11: phi3.block_count u32 = 40 llama_model_loader: - kv 12: phi3.attention.head_count u32 = 40 llama_model_loader: - kv 13: phi3.attention.head_count_kv u32 = 10 llama_model_loader: - kv 14: phi3.attention.layer_norm_rms_epsilon f32 = 0.000010 llama_model_loader: - kv 15: phi3.rope.dimension_count u32 = 128 llama_model_loader: - kv 16: phi3.rope.freq_base f32 = 250000.000000 llama_model_loader: - kv 17: general.file_type u32 = 18 llama_model_loader: - kv 18: phi3.attention.sliding_window u32 = 100352 llama_model_loader: - kv 19: tokenizer.ggml.model str = llama llama_model_loader: - kv 20: tokenizer.ggml.pre str = default llama_model_loader: - kv 21: tokenizer.ggml.tokens arr[str,100352] = ["", "▁Ġ", "er", "in", "on", ... llama_model_loader: - kv 22: tokenizer.ggml.scores arr[f32,100352] = [0.000000, -0.000000, -1.000000, -2.0... llama_model_loader: - kv 23: tokenizer.ggml.token_type arr[i32,100352] = [2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ... llama_model_loader: - kv 24: tokenizer.ggml.bos_token_id u32 = 100257 llama_model_loader: - kv 25: tokenizer.ggml.eos_token_id u32 = 100257 llama_model_loader: - kv 26: tokenizer.ggml.padding_token_id u32 = 100257 llama_model_loader: - kv 27: tokenizer.chat_template str = {% for message in messages %}{% if (m... llama_model_loader: - kv 28: general.quantization_version u32 = 2 llama_model_loader: - type f32: 81 tensors llama_model_loader: - type f16: 1 tensors llama_model_loader: - type q6_K: 161 tensors llm_load_vocab: SPM vocabulary, but newline token not found: unordered_map::at! Using special_pad_id instead.llm_load_vocab: special tokens cache size = 97 llm_load_vocab: token to piece cache size = 0.7072 MB llm_load_print_meta: format = GGUF V3 (latest) llm_load_print_meta: arch = phi3 llm_load_print_meta: vocab type = SPM llm_load_print_meta: n_vocab = 100352 llm_load_print_meta: n_merges = 0 llm_load_print_meta: vocab_only = 0 llm_load_print_meta: n_ctx_train = 16384 llm_load_print_meta: n_embd = 5120 llm_load_print_meta: n_layer = 40 llm_load_print_meta: n_head = 40 llm_load_print_meta: n_head_kv = 10 llm_load_print_meta: n_rot = 128 llm_load_print_meta: n_swa = 100352 llm_load_print_meta: n_embd_head_k = 128 llm_load_print_meta: n_embd_head_v = 128 llm_load_print_meta: n_gqa = 4 llm_load_print_meta: n_embd_k_gqa = 1280 llm_load_print_meta: n_embd_v_gqa = 1280 llm_load_print_meta: f_norm_eps = 0.0e+00 llm_load_print_meta: f_norm_rms_eps = 1.0e-05 llm_load_print_meta: f_clamp_kqv = 0.0e+00 llm_load_print_meta: f_max_alibi_bias = 0.0e+00 llm_load_print_meta: f_logit_scale = 0.0e+00 llm_load_print_meta: n_ff = 17920 llm_load_print_meta: n_expert = 0 llm_load_print_meta: n_expert_used = 0 llm_load_print_meta: causal attn = 1 llm_load_print_meta: pooling type = 0 llm_load_print_meta: rope type = 2 llm_load_print_meta: rope scaling = linear llm_load_print_meta: freq_base_train = 250000.0 llm_load_print_meta: freq_scale_train = 1 llm_load_print_meta: n_ctx_orig_yarn = 16384 llm_load_print_meta: rope_finetuned = unknown llm_load_print_meta: ssm_d_conv = 0 llm_load_print_meta: ssm_d_inner = 0 llm_load_print_meta: ssm_d_state = 0 llm_load_print_meta: ssm_dt_rank = 0 llm_load_print_meta: ssm_dt_b_c_rms = 0 llm_load_print_meta: model type = 14B llm_load_print_meta: model ftype = Q6_K llm_load_print_meta: model params = 14.66 B llm_load_print_meta: model size = 11.77 GiB (6.89 BPW) llm_load_print_meta: general.name = Phi 4 llm_load_print_meta: BOS token = 100257 '<|endoftext|>' llm_load_print_meta: EOS token = 100257 '<|endoftext|>' llm_load_print_meta: EOT token = 100265 '<|im_end|>' llm_load_print_meta: UNK token = 0 '' llm_load_print_meta: PAD token = 100257 '<|endoftext|>' llm_load_print_meta: FIM PRE token = 100258 '<|fim_prefix|>' llm_load_print_meta: FIM SUF token = 100260 '<|fim_suffix|>' llm_load_print_meta: FIM MID token = 100259 '<|fim_middle|>' llm_load_print_meta: EOG token = 100257 '<|endoftext|>' llm_load_print_meta: EOG token = 100265 '<|im_end|>' llm_load_print_meta: max token length = 33 llm_load_tensors: offloading 40 repeating layers to GPU llm_load_tensors: offloading output layer to GPU llm_load_tensors: offloaded 41/41 layers to GPU llm_load_tensors: CPU_Mapped model buffer size = 980.00 MiB llm_load_tensors: CUDA0 model buffer size = 5599.45 MiB llm_load_tensors: CUDA1 model buffer size = 5468.14 MiB ................................................................................... llama_new_context_with_model: n_seq_max = 1 llama_new_context_with_model: n_ctx = 2048 llama_new_context_with_model: n_ctx_per_seq = 2048 llama_new_context_with_model: n_batch = 2048 llama_new_context_with_model: n_ubatch = 512 llama_new_context_with_model: flash_attn = 0 llama_new_context_with_model: freq_base = 250000.0 llama_new_context_with_model: freq_scale = 1 llama_new_context_with_model: n_ctx_per_seq (2048) < n_ctx_train (16384) -- the full capacity of the model will not be utilized llama_kv_cache_init: CUDA0 KV buffer size = 210.00 MiB llama_kv_cache_init: CUDA1 KV buffer size = 190.00 MiB llama_new_context_with_model: KV self size = 400.00 MiB, K (f16): 200.00 MiB, V (f16): 200.00 MiB llama_new_context_with_model: CUDA_Host output buffer size = 0.38 MiB llama_new_context_with_model: pipeline parallelism enabled (n_copies=6) llama_new_context_with_model: CUDA0 compute buffer size = 289.01 MiB llama_new_context_with_model: CUDA1 compute buffer size = 310.02 MiB llama_new_context_with_model: CUDA_Host compute buffer size = 34.04 MiB llama_new_context_with_model: graph nodes = 1606 llama_new_context_with_model: graph splits = 3 common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable) main: llama threadpool init, n_threads = 8

sampler seed: 96750315 sampler params: repeat_last_n = 64, repeat_penalty = 1.000, frequency_penalty = 0.000, presence_penalty = 0.000 dry_multiplier = 0.000, dry_base = 1.750, dry_allowed_length = 2, dry_penalty_last_n = -1 top_k = 40, top_p = 0.950, min_p = 0.050, xtc_probability = 0.000, xtc_threshold = 0.100, typical_p = 1.000, temp = 0.800 mirostat = 0, mirostat_lr = 0.100, mirostat_ent = 5.000 sampler chain: logits -> logit-bias -> penalties -> dry -> top-k -> typical -> top-p -> min-p -> xtc -> temp-ext -> dist generate: n_ctx = 2048, n_batch = 2048, n_predict = 256, n_keep = 1

Tell me a joke.ordord Ġteaspoon ĠI Ġteaspoon g Ġteaspoon ĠV Ġteaspoon g Ġteaspoonart Ġteaspoon Ġk Ġteaspoon Aord Ġteaspoonill Ġteaspoon g Ġteaspoon i Ġteaspoonher Ġteaspoon g Ġv Ġtriplet Ġteaspoonart Ġteaspoon Ġk Ġteaspoon ĠIord Ġteaspoon i Ġteaspoon g Ġteaspoon ĠV Ġteaspoonher Ġteaspoon Ġk Ġteaspoonra Ġteaspoon , Ġteaspoon Ġk Ġteaspoon1⁄4ord Ġteaspoon Ġun Ġteaspoon Ġk ĠteaspoonRE Ġteaspoonher Ġteaspoon g Ġteaspoon , Ġteaspoon Ġkord Ġteaspoon1⁄4 Ġteaspoon A Ġteaspoon , Ġteaspoon Ġk Ġteaspoon Aord Ġteaspoon i Ġteaspoon g Ġteaspoonill Ġteaspoonell Ġteaspoon g Ġteaspoon ĠVord Ġteaspoon1⁄4 Ġteaspoonill Ġv Ġtriplet Ġteaspoon ĠD Ġteaspoon1⁄4 Ġteaspoon); Ġteaspoon1⁄4 Ġteaspoon Aord Ġteaspoonell Ġteaspoon1⁄4 Ġteaspoonher Ġteaspoon1⁄4 Ġteaspoonell Ġteaspoon ĠV Ġteaspoon Ġk Ġteaspoon); Ġv Ġtriplet Ġteaspoon ĠIord Ġteaspoonell Ġteaspoon1⁄4ord Ġteaspoon1⁄4 Ġteaspoon A Ġteaspoon , Ġteaspoon g Ġteaspoonart Ġteaspoon Ġk Ġteaspoon Aord Ġteaspoon); Ġv Ġtriplet Ġteaspoonill Ġteaspoon g ĠteaspoonRE Ġteaspoon g Ġteaspoonart Ġteaspoon Aord Ġteaspoon i Ġteaspoon g Ġteaspoonher Ġteaspoon A Ġteaspoon1⁄4 Ġteaspoonher Ġteaspoon , Ġv Ġtriplet Ġteaspoon ĠIord Ġteaspoon A Ġteaspoon1⁄4 Ġteaspoon Ġk Ġteaspoonell Ġteaspoon g Ġteaspoon); Ġteaspoonest Ġteaspoon Ġk Ġteaspoon Ġg Ġteaspoon Ġk Ġteaspoonct Ġteaspoon1⁄4 Ġteaspoon ĠD Ġteaspoon Ġk Ġv Ġtripletord ĠteaspoonRE Ġteaspoon Ġk Ġteaspoon ĠD Ġteaspoonop Ġteaspoonher Ġteaspoon g Ġteaspoonart Ġteaspoon Ġk Ġteaspoon ĠIar [end of text]

llama_perf_sampler_print: sampling time = 6.05 ms / 246 runs ( 0.02 ms per token, 40634.29 tokens per second) llama_perf_context_print: load time = 1693.08 ms llama_perf_context_print: prompt eval time = 26.42 ms / 7 tokens ( 3.77 ms per token, 264.96 tokens per second) llama_perf_context_print: eval time = 3993.62 ms / 238 runs ( 16.78 ms per token, 59.60 tokens per second) llama_perf_context_print: total time = 4034.65 ms / 245 tokens

MS model card follows

Phi-4 Model Card

Model Summary


Developers	Microsoft Research
Description	`phi-4` is a state-of-the-art open model built upon a blend of synthetic datasets, data from filtered public domain websites, and acquired academic books and Q&A datasets. The goal of this approach was to ensure that small capable models were trained with data focused on high quality and advanced reasoning. `phi-4` underwent a rigorous enhancement and alignment process, incorporating both supervised fine-tuning and direct preference optimization to ensure precise instruction adherence and robust safety measures
Architecture	14B parameters, dense decoder-only Transformer model
Inputs	Text, best suited for prompts in the chat format
Context length	16K tokens
GPUs	1920 H100-80G
Training time	21 days
Training data	9.8T tokens
Outputs	Generated text in response to input
Dates	October 2024 – November 2024
Status	Static model trained on an offline dataset with cutoff dates of June 2024 and earlier for publicly available data
Release date	December 12, 2024
License	MSRLA

Intended Use


Primary Use Cases	Our model is designed to accelerate research on language models, for use as a building block for generative AI powered features. It provides uses for general purpose AI systems and applications (primarily in English) which require: 1. Memory/compute constrained environments. 2. Latency bound scenarios. 3. Reasoning and logic.
Out-of-Scope Use Cases	Our models is not specifically designed or evaluated for all downstream purposes, thus: 1. Developers should consider common limitations of language models as they select use cases, and evaluate and mitigate for accuracy, safety, and fairness before using within a specific downstream use case, particularly for high-risk scenarios. 2. 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, including the model’s focus on English. 3. Nothing contained in this Model Card should be interpreted as or deemed a restriction or modification to the license the model is released under.

Primary Use Cases

Our model is designed to accelerate research on language models, for use as a building block for generative AI powered features. It provides uses for general purpose AI systems and applications (primarily in English) which require:

1. Memory/compute constrained environments.
2. Latency bound scenarios.
3. Reasoning and logic.

Out-of-Scope Use Cases

Our models is not specifically designed or evaluated for all downstream purposes, thus:

1. Developers should consider common limitations of language models as they select use cases, and evaluate and mitigate for accuracy, safety, and fairness before using within a specific downstream use case, particularly for high-risk scenarios.
2. 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, including the model’s focus on English.
3. Nothing contained in this Model Card should be interpreted as or deemed a restriction or modification to the license the model is released under.

Data Overview

Training Datasets

Our training data is an extension of the data used for Phi-3 and includes a wide variety of sources from:

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.).
Acquired academic books and Q&A datasets.
High quality chat format supervised data covering various topics to reflect human preferences on different aspects such as instruct-following, truthfulness, honesty and helpfulness.

Multilingual data constitutes about 8% of our overall data. 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.

Benchmark datasets

We evaluated phi-4 using OpenAI’s SimpleEval and our own internal benchmarks to understand the model’s capabilities, more specifically:

MMLU: Popular aggregated dataset for multitask language understanding.
MATH: Challenging competition math problems.
GPQA: Complex, graduate-level science questions.
DROP: Complex comprehension and reasoning.
MGSM: Multi-lingual grade-school math.
HumanEval: Functional code generation.
SimpleQA: Factual responses.

Safety

Approach

phi-4 has adopted a robust safety post-training approach. This approach leverages a variety of both open-source and in-house generated synthetic datasets. The overall technique employed to do the safety alignment is a combination of SFT (Supervised Fine-Tuning) and iterative DPO (Direct Preference Optimization), including publicly available datasets focusing on helpfulness and harmlessness as well as various questions and answers targeted to multiple safety categories.

Safety Evaluation and Red-Teaming

Prior to release, phi-4 followed a multi-faceted evaluation approach. Quantitative evaluation was conducted with multiple open-source safety benchmarks and in-house tools utilizing adversarial conversation simulation. For qualitative safety evaluation, we collaborated with the independent AI Red Team (AIRT) at Microsoft to assess safety risks posed by phi-4 in both average and adversarial user scenarios. In the average user scenario, AIRT emulated typical single-turn and multi-turn interactions to identify potentially risky behaviors. The adversarial user scenario tested a wide range of techniques aimed at intentionally subverting the model’s safety training including jailbreaks, encoding-based attacks, multi-turn attacks, and adversarial suffix attacks.

Please refer to the technical report for more details on safety alignment.

Model Quality

To understand the capabilities, we compare phi-4 with a set of models over OpenAI’s SimpleEval benchmark.

At the high-level overview of the model quality on representative benchmarks. For the table below, higher numbers indicate better performance:

Category	Benchmark	phi-4 (14B)	phi-3 (14B)	Qwen 2.5 (14B instruct)	GPT-4o-mini	Llama-3.3 (70B instruct)	Qwen 2.5 (72B instruct)	GPT-4o
Popular Aggregated Benchmark	MMLU	84.8	77.9	79.9	81.8	86.3	85.3	88.1
Science	GPQA	56.1	31.2	42.9	40.9	49.1	49.0	50.6
Math	MGSM MATH	80.6 80.4	53.5 44.6	79.6 75.6	86.5 73.0	89.1 66.3*	87.3 80.0	90.4 74.6
Code Generation	HumanEval	82.6	67.8	72.1	86.2	78.9*	80.4	90.6
Factual Knowledge	SimpleQA	3.0	7.6	5.4	9.9	20.9	10.2	39.4
Reasoning	DROP	75.5	68.3	85.5	79.3	90.2	76.7	80.9

* These scores are lower than those reported by Meta, perhaps because simple-evals has a strict formatting requirement that Llama models have particular trouble following. We use the simple-evals framework because it is reproducible, but Meta reports 77 for MATH and 88 for HumanEval on Llama-3.3-70B.

Usage

Input Formats

Given the nature of the training data, phi-4 is best suited for prompts using the chat format as follows:

<|im_start|>system<|im_sep|>
You are a medieval knight and must provide explanations to modern people.<|im_end|>
<|im_start|>user<|im_sep|>
How should I explain the Internet?<|im_end|>
<|im_start|>assistant<|im_sep|>

With `transformers`

import transformers

pipeline = transformers.pipeline(
    "text-generation",
    model="microsoft/phi-4",
    model_kwargs={"torch_dtype": "auto"},
    device_map="auto",
)

messages = [
    {"role": "system", "content": "You are a medieval knight and must provide explanations to modern people."},
    {"role": "user", "content": "How should I explain the Internet?"},
]

outputs = pipeline(messages, max_new_tokens=128)
print(outputs[0]["generated_text"][-1])

Responsible AI Considerations

Like other language models, phi-4 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 model is trained primarily on English 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. phi-4 is not intended to support multilingual use.
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-4 training data is based in Python and uses 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.). Using safety services like Azure AI Content Safety that have advanced guardrails is highly recommended. 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 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.