9B Model Training with Bidirectional Refinement Loops
Train a 9B model with a reverse LLM sidecar to improve syntax accuracy and run full HumanEval.
Run full HumanEval on both models to compare performance.
Summary
A developer trains a 9B model and plans to run the full HumanEval dataset, building on a 1.7B model that can generate code. The approach uses a reverse LLM sidecar that reads from the end of the output and injects its result back at the top, creating a bidirectional refinement loop focused on syntax improvement.
The training is inspired by neuroanatomy findings and aims to drastically improve a very small model's code generation capabilities. The author intends to run the full HumanEval dataset on both the 1.7B and 9B models, rather than just the first 20 samples.
The post outlines the methodology and future plans for evaluation.
Key changes
- Training 9B model
- Uses reverse LLM sidecar for bidirectional refinement
- Focuses on syntax improvement
- Plans to run full HumanEval dataset
- Inspired by neuroanatomy findings
- Model size 1.7B for code generation