keras/phi3_mini_4k_instruct_en

Model Overview

Phi-3 is a set of large language models published by Microsoft. Models are instruction tuned, and range in size from 3 billion to 14 billion parameters. See the model card below for benchmarks, data sources, and intended use cases.

Weights are released under the MIT License. Keras model code is released under the Apache 2 License.

Links

• Phi-3 Quickstart Notebook
• Phi-3 API Documentation
• Phi-3 Model Card
• KerasHub Beginner Guide
• KerasHub Model Publishing Guide

Installation

Keras and KerasHub can be installed with:

pip install -U -q keras-hub
pip install -U -q keras>=3

Jax, TensorFlow, and Torch come preinstalled in Kaggle Notebooks. For instruction on installing them in another environment see the Keras Getting Started page.

Presets

The following model checkpoints are provided by the Keras team. Full code examples for each are available below.

Preset name	Parameters	Description
phi3_mini_4k_instruct_en	3.82B	3B model with 4K max context
phi3_mini_128k_instruct_en	3.82B	3B model with 128K max context

Prompts

Phi-3 models are instruction tuned on turn by turn conversations and should be prompted with examples that precisely match the training data. Specifically, you must alternate user and assistant turns that begin and end with special tokens. New lines do matter. See the following for an example:

prompt = """<|user|>
Hello!<|end|>
<|assistant|>
Hello! How are you?<|end|>
<|user|>
I'm great. Could you help me with a task?<|end|>
"""

Example Usage

pip install -U -q keras-hub

import keras
import keras_hub
import numpy as np

Use generate() to do text generation.

phi3_lm = keras_hub.models.Phi3CausalLM.from_preset("phi3_mini_4k_instruct_en")
phi3_lm.generate("<|user|>\nHow to explain Internet for a medieval knight?<|end|>\n<|assistant|>", max_length=500)

# Generate with batched prompts.
phi3_lm.generate([
    "<|user|>\nWhat is Keras?<|end|>\n<|assistant|>",
    "<|user|>\nGive me your best brownie recipe.<|end|>\n<|assistant|>",
], max_length=500)

Compile the generate() function with a custom sampler.

phi3_lm = keras_hub.models.Phi3CausalLM.from_preset("phi3_mini_4k_instruct_en")
phi3_lm.compile(sampler="greedy")
phi3_lm.generate("<|user|>\nWhat is Keras?<|end|>\n<|assistant|>", max_length=30)

phi3_lm.compile(sampler=keras_hub.samplers.BeamSampler(num_beams=2))
phi3_lm.generate("<|user|>\nWhat is Keras?<|end|>\n<|assistant|>", max_length=30)

Use generate() without preprocessing.

prompt = {
    "token_ids": np.array([[306, 864, 304, 1827, 0, 0, 0, 0, 0, 0]] * 2),
    # Use `"padding_mask"` to indicate values that should not be overridden.
    "padding_mask": np.array([[1, 1, 1, 1, 0, 0, 0, 0, 0, 0]] * 2),
}

phi3_lm = keras_hub.models.Phi3CausalLM.from_preset(
    "phi3_mini_4k_instruct_en",
    preprocessor=None,
    dtype="bfloat16"
)
phi3_lm.generate(prompt)

Call fit() on a single batch.

features = ["The quick brown fox jumped.", "I forgot my homework."]
phi3_lm = keras_hub.models.Phi3CausalLM.from_preset("phi3_mini_4k_instruct_en")
phi3_lm.fit(x=features, batch_size=2)

Example Usage with Hugging Face URI

pip install -U -q keras-hub

import keras
import keras_hub
import numpy as np

Use generate() to do text generation.

phi3_lm = keras_hub.models.Phi3CausalLM.from_preset("hf://keras/phi3_mini_4k_instruct_en")
phi3_lm.generate("<|user|>\nHow to explain Internet for a medieval knight?<|end|>\n<|assistant|>", max_length=500)

# Generate with batched prompts.
phi3_lm.generate([
    "<|user|>\nWhat is Keras?<|end|>\n<|assistant|>",
    "<|user|>\nGive me your best brownie recipe.<|end|>\n<|assistant|>",
], max_length=500)

Compile the generate() function with a custom sampler.

phi3_lm = keras_hub.models.Phi3CausalLM.from_preset("hf://keras/phi3_mini_4k_instruct_en")
phi3_lm.compile(sampler="greedy")
phi3_lm.generate("<|user|>\nWhat is Keras?<|end|>\n<|assistant|>", max_length=30)

phi3_lm.compile(sampler=keras_hub.samplers.BeamSampler(num_beams=2))
phi3_lm.generate("<|user|>\nWhat is Keras?<|end|>\n<|assistant|>", max_length=30)

Use generate() without preprocessing.

prompt = {
    "token_ids": np.array([[306, 864, 304, 1827, 0, 0, 0, 0, 0, 0]] * 2),
    # Use `"padding_mask"` to indicate values that should not be overridden.
    "padding_mask": np.array([[1, 1, 1, 1, 0, 0, 0, 0, 0, 0]] * 2),
}

phi3_lm = keras_hub.models.Phi3CausalLM.from_preset(
    "hf://keras/phi3_mini_4k_instruct_en",
    preprocessor=None,
    dtype="bfloat16"
)
phi3_lm.generate(prompt)

Call fit() on a single batch.

features = ["The quick brown fox jumped.", "I forgot my homework."]
phi3_lm = keras_hub.models.Phi3CausalLM.from_preset("hf://keras/phi3_mini_4k_instruct_en")
phi3_lm.fit(x=features, batch_size=2)