auryn-macmillan/boostedv1-ilv9
The auryn-macmillan/boostedv1-ilv9 model is a 1.5 billion parameter language model, an Iterative-Learning (IL v9) version of BoostedV1, built upon deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B. It was developed by auryn-macmillan through continued LoRA training and three rounds of self-improvement. This model demonstrates notable gains in code generation, specifically on the HumanEval benchmark, making it suitable for programming-related tasks.
Loading preview...
Model Overview
auryn-macmillan/boostedv1-ilv9 is a 1.5 billion parameter language model, representing the Iterative-Learning (IL v9) iteration of BoostedV1. It is based on the deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B architecture and was enhanced through a multi-round self-improvement process.
Key Capabilities and Training
This model was developed by continuing LoRA training from the BoostedV1 Phase-1 model, followed by three rounds of an IL v9 self-improvement loop. Each round involved generating candidate solutions, external verification, model self-auditing, and subsequent training on high-quality filtered data. The most significant improvements were observed in code generation capabilities.
Performance Highlights
Evaluation results indicate a notable improvement in code generation performance:
- HumanEval (pass@1): Increased from 7.3% (BoostedV1 Phase-1) to 11.0% (BoostedV1-ILv9).
- GSM8K: Showed a slight improvement from 46.0% to 46.5%.
These gains highlight the effectiveness of the IL v9 self-improvement process, particularly for programming tasks. The model utilizes a standard Qwen2 architecture and does not require custom code or trust_remote_code.
Ideal Use Cases
- Code Generation: Particularly strong in generating Python code, as evidenced by its improved HumanEval scores.
- Programming Assistance: Suitable for tasks requiring code completion or basic code problem-solving.
- Research in Iterative Learning: Can serve as a case study for models enhanced through self-improvement loops.