This is my take on implementing the Self-Instruct paper. The approach is quite heavily modified, and does not use any human-generated seeds.
This updated implementation supports either the /v1/completions endpoint or /v1/chat/completions, which is particularly useful in that it supports gpt-4 and gpt-3.5-turbo (which is 1/10 the cost of text-davinci-003).
Huge thank you to the folks over at a16z for sponsoring the costs associated with building models and associated tools!
via pip:
pip install --no-build-isolation airoboros
from source (keeping the source):
git clone https://github.com/jondurbin/airoboros
pip install -e --no-build-isolation ./airoboros
- support for either /v1/completions or /v1/chat/completions APIs (which allows gpt-3.5-turbo instead of text-davinci-003, as well as gpt-4 if you have access)
- support for custom topics list, custom topic generation prompt, or completely random topics
- in-memory vector db (Chroma) for similarity comparison, which is much faster than calculating rouge score for each generated instruction
- (seemingly) better prompts, which includes injection of random topics to relate the instructions to, which creates much more diverse synthetic instructions
- asyncio producers with configurable batch size
- several "instructors", each targetting specific use-cases, such as Orca style reasoning/math, role playing, etc.
- tries to ensure the context, if provided, is relevant to the topic and contains all the information that would be necessary to respond to the instruction, and nost just a link to article/etc.
- generally speaking, this implementation tries to reduce some of the noise
Problem and proposed solution:
- Models can only ever be as good as the data they are trained on.
- High quality data is difficult to curate manually, so ideally the process can be automated by AI/LLMs.
- Large models (gpt-4, etc.) are pricey to build/run and out of reach for individuals/small-medium business, and are subject to RLHF bias, censorship, and changes without notice.
- Smaller models (llama-2-70b, etc.) can reach somewhat comparable performance in specific tasks to much larger models when trained on high quality data.
- The airoboros tool allows building datasets that are focused on specific tasks, which can then be used to build a plethora of individual expert models. This means we can crowdsource building experts.
- Using either a classifier model, or simply calculating vector embeddings for each item in the dataset and using faiss index/cosine similarity/etc. search, incoming requests can be routed to a particular expert (e.g. dynamically loading LoRAs) to get extremely high quality responses.
Progress:
- ✅ PoC that training via self-instruction, that is, datasets generated from language models, works reasonably well.
- ✅ Iterate on the PoC to use higher quality prompts, more variety of instructions, etc.
- ✅ Split the code into separate "instructors", for specializing in any particular task (creative writing, songs, roleplay, coding, execution planning, function calling, etc.)
- [in progress]: PoC that an ensemble of LoRAs split by the category (i.e., the instructor used in airoboros) has better performance than the same param count model tuned on all data
- [in progress]: Remove the dependency on OpenAI/gpt-4 to generate the training data so all datasets can be completely free and open source.
- [future]: Automatic splitting of experts at some threshold, e.g. "coding" is split into python, js, golang, etc.
- [future]: Hosted service/site to build and/or extend datasets or models using airoboros.
- [future]: Depending on success of all of the above, potentially a hosted inference option with an exchange for private/paid LoRAs.
LMoE is the simplest architecture I can think of for a mixture of experts. It doesn't use a switch transformer, doesn't require slicing and merging layers with additional fine-tuning, etc. It just dynamically loads the best PEFT/LoRA adapter model based on the incoming request.
By using this method, we can theoretically crowdsource generation of dozens (or hundreds/thousands?) of very task-specific adapters and have an extremely powerful ensemble of models with very limited resources on top of a single base model (llama-2 7b/13b/70b).
The self-instruct code contained within this project uses many different "instructors" to generate training data to accomplish specific tasks. The output includes the instructor/category that generated the data. We can use this to automatically segment the training data to fine-tune specific "experts".
See scripts/segment_experts.py for an example of how the training data can be segmented, with a sampling of each other expert in the event of misrouting.
See scripts/tune_expert.py for an example of creating the adapter models (with positional args for expert name, model size, etc.)
NOTE: this assumes use of my fork of qlora https://github.com/jondurbin/qlora
The "best" routing mechanism would probably be to train a classifier based on the instructions for each category, with the category/expert being the label, but that prohibits dynamic loading of new experts.
Instead, this supports 3 options:
- faiss index similarity search using the training data for each expert (default)
- agent-based router using the "function" expert (query the LLM with a list of available experts and their descriptions, ask which would be best based on the user's input)
- specify the agent in the JSON request
First, download the base llama-2 model for whichever model size you want, e.g.: llama-2-7b-hf
Next, download the LMoE package that corresponds to that base model, e.g.: airoboros-lmoe-7b-2.1
NOTE: 13b also available, 70b in progress
Here's an example command to start the server:
python -m airoboros.lmoe.api \
--base-model ./llama-2-7b-hf \
--lmoe ./airoboros-lmoe-7b-2.1 \
--router-max-samples 1000 \
--router-k 25 \
--port 8000 \
--host 127.0.0.1
to use the agent-based router, add --agent-router to the arguments
This uses flash attention via bettertransformers (in optimum). You may need to install torch nightly if you see an error like 'no kernel available', e.g.:
pip install -U --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu118
Once started, you can infer using the same API scheme you'd query OpenAI API with, e.g.:
curl -H 'content-type: application/json' http://127.0.0.1:8000/v1/chat/completions -d '
{
"model": "llama-2-7b-hf",
"temperature": 0.7,
"max_tokens": 2048,
"messages": [
{
"role": "system",
"content": "A chat."
},
{
"role": "user",
"content": "How much wood would a woodchuck chuck if a woodchuck could chuck wood?"
}
]
}'
I've also added an vllm-based server, but the results aren't quite as good (not sure why yet). To use it, make sure you install vllm and fschat, or pip install airoboros[vllm]
python -m airoboros.lmoe.vllm \
--model ./llama-2-7b-hf \
--lmoe-path ../airoboros-lmoe-7b-2.1 \
--router-max-samples 100 \
--router-k 25 \
--port 8000 \
--host 127.0.0.1
To better accommodate the plethora of options, the configuration has been moved to a YAML config file.
Please create a copy of example-config.yaml and configure as desired.
Once you have the desired configuration, run:
airoboros generate-instructions --config-path /path/to/config.yaml
Again, this is now all YAML configuration based! Please create a customized version of the YAML config file, then run:
airoboros generate-topics --config-path /path/to/config.yaml
You can override the topic_prompt string in the configuration to use a different topic generation prompt.
ETH 0xce914eAFC2fe52FdceE59565Dd92c06f776fcb11
BTC bc1qdwuth4vlg8x37ggntlxu5cjfwgmdy5zaa7pswf
2.1 dataset
2.0/m2.0
- airoboros-l2-7b-gpt4-2.0
- airoboros-l2-7b-gpt4-m2.0
- airoboros-l2-13b-gpt4-2.0
- airoboros-l2-13b-gpt4-m2.0
Previous generation (1.4.1 dataset)
Latest version (2.0 / m2.0 datasets)
Previous generation (1.4.1 dataset)
- airoboros-65b-gpt4-1.4
- airoboros-33b-gpt4-1.4
- airoboros-13b-gpt4-1.4
- airoboros-7b-gpt4-1.4
- older versions on HF as well
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent