Custom Ladder implementation on any Complex problem for LLM. a reimplementation of the paper “LADDER: SELF-IMPROVING LLMS THROUGH RECURSIVE PROBLEM DECOMPOSITION” https://arxiv.org/pdf/2503.00735
git clone [email protected]:AbdelrahmanAbounida/ladder.git
cd ladder
pdm installpython src/main.py- LLM Intelligence ratio Equation
- Custom Verification Method if required (for our Graph Usecase)
- DatasetGenerator > Generate subproblems according to the model intelligence ratio (step3)
- Difficulty Engine should decide the level of difficulty to be generated and what transformations to be applied
- Verification engine should use the small llm to be tuned not the Larger one
- LLM Engine (temperature cycling and persona based prompts for different operations like variant generation)
- Ladder Finetuning Process
- GRPO Implementation
- reward functions
- TTRL Implementation
- Data Generation in a loop
- General Configurations for all Constants and Hyper Parameters
- implement different interfaced for different models to be used (HF, Ollama, VLLM, deepspeed, LiteLLM,..)
- LLMS Benchmarking
- Metrics and other evaluation methods
- implement more usecases if required for diverse benchmarking
- use accelerate / PEFT / deepspeed and vllm to speed up training process
- Documentation
- packaging
- CICD
- Testing