autologic is a Python package that implements the SELF-DISCOVER framework proposed in the paper "SELF-DISCOVER: Large Language Models Self-Compose Reasoning Structures".
It provides a way for large language models (LLMs) to automatically compose modular reasoning structures to tackle complex reasoning tasks, without the need for training data or labels.
- Implements the full SELF-DISCOVER pipeline enabling LLMs to self-discover reasoning structures
- Works with Gemini Pro and local GGUF models via llama-cpp-python
- Flexible integration powered by a simple LLMConfig
- Interactive prompts or standalone execution
- CLI and Python API access
The SELF-DISCOVER framework consists of two key stages:
Stage 1: Self-discover a reasoning structure for the task from a set of seed "reasoning modules"
Stage 2: Solve instances by following the composed structure
The first stage has 3 steps guided by meta-prompting:
- SELECT relevant reasoning modules
- ADAPT modules to be task-specific
- IMPLEMENT structure with adapted modules
The following instructions show the CMAKE arguments for compiling llama-cpp-python dependencies with support for metal. Refer to https://github.com/abetlen/llama-cpp-python for instructions regarding different GPUs.
git clone https://github.com/waszumteufel/autologic.git
cd autologic
python3 -m venv venv
source venv/bin/activate
CMAKE_ARGS="-DLLAMA_METAL=on" pip3 install -e .python3 -m venv venv
source venv/bin/activate
CMAKE_ARGS="-DLLAMA_METAL=on" pip3 install git+https://github.com/waszumteufel/autologic@main#egg=autologicTo use the Gemini model, set the GEMINI_PRO_API_KEY variable:
GEMINI_PRO_API_KEY="sk-..."
If this is not set, solve() will automatically read it from the .env file in the working directory or general environment.
It is recommended to set sensitive values like API keys in a .env file.
This will be automatically loaded from the current working directory if present, and used to populate any unset configuration values that have matching environment variable names (i.e. GEMINI_PRO_API_KEY).
See the "Configuration" section of the documentation for the full set of options. Using .env and environment variables allows keeping credentials secure.
This allows conveniently configuring API keys and other settings without exposing them directly in code. The LLMConfig can then be initialized without explicitly passing these values every time.
The below code block illustrates how to use the Python API to use SELF-DISCOVER with a Mixtral GGUF.
from autologic import reasoningEngine
my_config = reasoningEngine.LLMConfig(
gguf_path="/tmp/mixtral-8x7b-instruct-v0.1.Q8_0.gguf",
context_length=8000,
model_type=reasoningEngine.ModelType.LOCAL,
temp=0.2,
chat_template=reasoningEngine.ChatTemplate.MIXTRAL_INSTRUCT,
threads=12
)
result = reasoningEngine.solve(task = "What is 2 + 2?", llmConfig=my_config)
print(result)The API Key for Gemini Pro is read from the GEMINI_PRO_API_KEY environment variable or .env file. It can optionally be passed in to through the reasoningModuleLLMConfig.api_key field.
from autologic import reasoningEngine
my_config = reasoningEngine.LLMConfig(
context_length=8000,
model_type=reasoningEngine.ModelType.GEMINI,
temp=0.2,
)
result = reasoningEngine.solve(task = "What is 2 + 2?", llmConfig=my_config)
print(result)Example with a prompt
autologic gemini --temp 0.2 \
--retries 3 \
--prompt "What weighs more? A pound of feathers or a pound of lead?"Example output
Thinking...
2024-02-14 18:50:34.054248 | Starting SELECT Phase
2024-02-14 18:50:37.008261 | SELECT Phase Complete
2024-02-14 18:50:37.008367 | Starting ADAPT Phase
2024-02-14 18:50:40.322604 | ADAPT Phase Complete
2024-02-14 18:50:40.322641 | Starting IMPLEMENT Phase
2024-02-14 18:50:45.567405 | IMPLEMENT Phase Complete
2024-02-14 18:50:45.567738 | Starting to Solve Problem using Reasoning Structure
2024-02-14 18:50:52.775653 | Solution has been found.
ANSWER: The weight of a pound of feathers is equal to the weight of a pound of lead.
The below example shows how to specify a path to a GGUF model and how to specify the prompt format for the given model with the --format flag.
The example also showcases the usage of the --verbose flag which shows detailed reasoning information as well as the reasoning structure that the LLM self-discovers for the given problem.
autologic local --temp 0.2 \
--retries 5 \
--prompt "What weighs more? A pound of feathers or a pound of lead?" \
--gguf_path /tmp/qwen1_5-72b-chat-q5_k_m.gguf \
--threads 12 \
--layers -1 \
--format chatml \
--verbose Example output:
Thinking...
2024-02-14 20:05:23.116629 | Starting SELECT Phase
2024-02-14 20:06:41.595659 | SELECT Phase Complete
2024-02-14 20:06:41.595711 | Reasoning Modules Picked:
- 3. How can I simplify the problem so that it is easier to solve?
- 9. Critical Thinking: This style involves analyzing the problem from different perspectives, questioning assumptions, and evaluating the evidence or information available. It focuses on logical reasoning, evidence-based decision-making, and identifying potential biases or flaws in thinking.
- 16. What are the underlying causes or factors contributing to the problem?
- 22. How can progress or success in solving the problem be measured or evaluated?
- 23. What indicators or metrics can be used?
2024-02-14 20:06:41.595719 | Starting ADAPT Phase
2024-02-14 20:07:22.370683 | ADAPT Phase Complete
2024-02-14 20:07:22.370713 | Task-specific Reasoning Module verbiage:
- Clarify the units and equalities: Both items are measured in pounds, which means they weigh the same amount by definition; a pound is a unit of weight or mass.
- Understand the context and potential misconceptions: The question is a classic example of a trick question, as some may think lead is denser and therefore heavier, but in reality, both weigh the same.
- Address the physical properties without affecting the weight: While lead is denser than feathers, density does not affect the weight for the same volume; both weigh one pound, regardless of their density.
2024-02-14 20:07:22.370720 | Starting IMPLEMENT Phase
2024-02-14 20:08:46.753718 | IMPLEMENT Phase Complete
2024-02-14 20:08:46.753805 | Reasoning Structure:
{
"Reasoning Structure": {
"Step 1: Understand the Weighing Units": {
"Description": "Both items are measured in the same unit.",
"Action": "Confirm that both 'pound of feathers' and 'pound of lead' use the same unit (pound).",
"Unit": "pound"
},
"Step 2: Recognize the Trick Question": {
"Description": "The question is designed to prompt a misconception.",
"Action": "Identify that some might think density affects weight, but it doesn't for the same volume.",
"Misconception": "Density affects weight"
},
"Step 3: Define Density": {
"Description": "Density is mass per unit volume.",
"Action": "Explain that density does not change the weight for the same amount of mass.",
"Density": "mass / volume"
},
"Step 4: Compare Weights Without Density": {
"Description": "Focus on the weight aspect.",
"Action": "Since both weigh one pound, compare them without considering density.",
"Comparison": "Equal weight"
},
"Step 5: Conclusion": {
"Description": "Both weigh the same amount.",
"Action": "Both a pound of feathers and a pound of lead weigh the same.",
"Final Conclusion": ""
},
"FINAL_ANSWER": ""
}
}
2024-02-14 20:08:46.753961 | Starting to Solve Problem using Reasoning Structure
2024-02-14 20:10:07.283210 | Problem Solved
Completed Reasoning Structure:
{
"Reasoning Structure": {
"Step 1: Understand the Weighing Units": {
"Description": "Both items are measured in the same unit.",
"Action": "Confirm that both 'pound of feathers' and 'pound of lead' use the same unit (pound).",
"Unit": "pound"
},
"Step 2: Recognize the Trick Question": {
"Description": "The question is designed to prompt a misconception.",
"Action": "Identify that some might think density affects weight, but it doesn't for the same volume.",
"Misconception": "Density affects weight"
},
"Step 3: Define Density": {
"Description": "Density is mass per unit volume.",
"Action": "Explain that density does not change the weight for the same amount of mass.",
"Density": "mass / volume"
},
"Step 4: Compare Weights Without Density": {
"Description": "Focus on the weight aspect.",
"Action": "Since both weigh one pound, compare them without considering density.",
"Comparison": "Equal weight"
},
"Step 5: Conclusion": {
"Description": "Both weigh the same amount.",
"Action": "Both a pound of feathers and a pound of lead weigh the same.",
"Final Conclusion": "A pound of feathers and a pound of lead weigh equally."
},
"FINAL_ANSWER": "A pound of feathers and a pound of lead weigh equally."
}
}
2024-02-14 20:10:07.283302 | Solution has been found.
ANSWER: A pound of feathers and a pound of lead weigh equally.
The CLI can be invoked without passing a --prompt argument to enter an interactive mode. This allows conveniently sending multi-line prompts for the model to reason over:
% autologic gemini --temp 0.2
Entering Interactive Mode: CTRL + C to send multi-line input. CTRL + D to exit the program.
> After starting interactive mode, you will be prompted to enter text:
Entering Interactive Mode: CTRL + C to send multi-line input. CTRL + D to exit the program.
> What weighs more?
> A pound of feathers?
> Or a pound of lead?
> Enter your reasoning prompt spanning multiple lines.
To submit the prompt, press CTRL + C to send the input while staying in interactive mode.
This will trigger the model to start reasoning over the prompt:
> What weighs more?
> a pound of feathers or
> a pound of lead?
> ^C
Thinking...
2024-02-14 20:17:38.307588 | Starting SELECT Phase
2024-02-14 20:17:42.403366 | SELECT Phase Complete
2024-02-14 20:17:42.403473 | Starting ADAPT Phase
2024-02-14 20:17:45.680310 | ADAPT Phase Complete
2024-02-14 20:17:45.680418 | Starting IMPLEMENT Phase
2024-02-14 20:17:52.029561 | IMPLEMENT Phase Complete
2024-02-14 20:17:52.029897 | Starting to Solve Problem using Reasoning Structure
2024-02-14 20:17:57.064476 | Solution has been found.
ANSWER: A pound of feathers weighs the same as a pound of lead.
Once you are done, send the CTRL + D signal to exit interactive mode and quit the CLI program.
> Goodbye!
This interactive workflow allows you to conveniently test long, complex reasoning without having to put all the prompt in quotes or escape newlines.
- Expose information Reasoning Structure and Reasoning Module selection via Python API. Currently, it is only visible when using verbose=True in CLI and API.
- Add Mixed mode where discovery of Reasoning Structure and usage of Reasoning Structure can be done by different LLMs.
- Add support for openai models
- Add package to Pypi
- Add support for other prompt formats (Llama2, airoboros, etc.)
- Add REST support via Flask