Caikit-NLP is a python library providing various Natural Language Processing (NLP) capabilities built on top of caikit framework.

Caikit-NLP implements concept of "task" from caikit framework to define (and consume) interfaces for various NLP problems and implements various "modules" to provide functionalities for these "modules".

Capabilities provided by caikit-nlp:

To help you quickly get started with using Caikit, we have prepared a Jupyter notebook that can be run in Google Colab. Caikit-nlp is a powerful library that leverages prompt tuning and fine-tuning to add NLP domain capabilities to caikit.

To install from git repo:

python -m venv .venv
source .venv/bin/activate
pip install git+https://github.com/caikit/caikit-nlp

caikit_nlp can use Hugging Face models, allowing for direct download and bootstrapping.

For example, to use google/flan-t5-small:

import os
# The env var ALLOW_DOWNLOADS has to be set to allow model downloads before importing caikit_nlp
os.environ['ALLOW_DOWNLOADS'] = "1"

import caikit_nlp

model_name = "google/flan-t5-small"
model = caikit_nlp.text_generation.TextGeneration.bootstrap(model_name)
model.save(f"{model_name}-caikit") # optionally save the model

To serve models, the following basic configuration can be used:

# config.yml
runtime:
  library: caikit_nlp
  local_models_dir: ./models

log:
  formatter: pretty # optional: log formatter is set to json by default

Start the server:

env CONFIG_FILES=./config.yml python -m caikit.runtime

The model can now be queried at localhost:8080 via http or at localhost:8085 via grpc.

For example, using the http server and using curl to send a POST request:

curl --json '{
    "model_id": "flan-t5-small-caikit",
    "inputs": "At what temperature does liquid Nitrogen boil?"
}' localhost:8080/api/v1/task/text-generation

We get the following response:

{
  "generated_text": "74 degrees F",
  "generated_tokens": 5,
  "finish_reason": "MAX_TOKENS",
  "producer_id": {
    "name": "Text Generation",
    "version": "0.1.0"
  },
  "input_token_count": 10,
  "seed": null
}

All the available API endpoints and protos can be dumped using scripts/dump_apis.sh.

To build the docker image:

python -m build --wheel
docker build -t caikit-nlp:latest .

A volume can be mounted at /caikit providing configuration and (optionally) models:

mkdir -p caikit
$EDITOR caikit/config.yml # edit as required
cp -r <path/to/models> ./caikit/models
docker run -e CONFIG_FILES=/caikit/config.yml -v $PWD/caikit/:/caikit -p 8080:8080 -p 8085:8085 python -m caikit.runtime

In order to start the serving runtime:

docker run -e CONFIG_FILES=/caikit/config.yml \
    -v $PWD/caikit/:/caikit -p 8080:8080 -p 8085 \
    python -m caikit.runtime

Assuming the standard configuration with port 8080 for the http server and 8085 for the grpc server.

Configuration can be provided via environment variables or by providing a yaml configuration file thanks to alchemy-config.

For example, to set the caikit runtime, setting RUNTIME_LIBRARY=caikit_nlp via environment variables or providing the following yaml configuration is equivalent.

# config.yml
runtime:
  library: caikit_nlp

For configuration options see caikit_nlp's example config: config.yml or caikit's example caikit.yml.

We welcome contributions from the community! If you would like to contribute to caikit-nlp, please read the guidelines in the main project's CONTRIBUTING.md file. It includes information on submitting bug reports, feature requests, and pull requests. Make sure to follow our coding standards, code of conduct, security standards, and documentation guidelines to streamline the contribution process.

This project is licensed under the ASFv2 License.

A list of terms that either may be unfamiliar or that have nebulous definitions based on who and where you hear them, defined for how they are used/thought of in the caikit/caikit-nlp project:

• Fine tuning - trains the base model onto new data etc; this changes the base model.
• Prompt engineering - (usually) manually crafting texts that make models do a better job that's left appended to the input text. E.g., if you wanted to do something like sentiment on movie reviews, you might come up with a prompt like The movie was: _____ and replace the _____ with the movie review you're consider to try to get something like happy/sad out of it.
• PEFT - library by Huggingface containing implementations of different tuning methods that scale well - things like prompt tuning, and MPT live there. So PEFT itself isn't an approach even though parameter efficient fine-tuning sounds like one. Prompt tuning - learning soft prompts. This is different from prompt engineering in that you're not trying to learn tokens. Instead, you're basically trying to learn new embedded representations (sometimes called virtual tokens) that can be concatenated onto your embedded input text to improve the performance. This can work well, but also can be sensitive to initialization.
• Multitask prompt tuning (MPT) - Tries to fix some of the issues with prompt tuning by allowing you to effectively learn 'source prompts' across different tasks & leverage them to initialize your prompt tuning etc. More information on MPT can be found at: https://arxiv.org/abs/2303.02861

The important difference between fine tuning and capabilities like prompt tuning/multi-taskprompt tuning is that the latter doesn't change the base model's weights at all. So when you run inference for prompt tuned models, you can have n prompts to 1 base model, and just inject the prompt tensors you need when they're requested instead of having n separate fine-tuned models.

Runtime Performance Benchmarking for tuning various models.

• Currently causal language models and sequence-to-sequence models are supported.