slvDev
Featured
esp32-ai
A 28.9 million parameter language model runs on an $8 ESP32-S3 microcontroller entirely on-device, generating simple stories at about 9.5 tokens per second.
AI repository intelligence
FR-AI / ANALYSISWhy it is worth attention
It introduces a structured pipeline for instilling self-correction capabilities in LLMs through synthetic preference pairs and explicit trajectory alignment, with clear evaluation metrics for correction success and over-revision rates.
Who it is for
- NLP researchers studying self-correction and reasoning in LLMs
- ML engineers building robust math reasoning systems
- Practitioners applying DPO and SFT for preference alignment
- Developers interested in automated data synthesis for training
Use cases
- Improving a model's ability to correct its own mathematical answers
- Building a system that automatically revises draft responses
- Reducing over-revision errors when the original draft is already correct
- Training a model to generate high-quality critiques along with corrected answers
Strengths
- Complete three-stage training recipe with ready-to-use scripts and configs
- Automatic synthetic preference pair generation from GSM8K, including repair and keep-correct types
- Supports efficient QLoRA / LoRA training with as little as 16GB GPU memory
- Comprehensive evaluation suite for self-correction (success rate, over-revision rate) and direct math accuracy
Considerations
- Synthetic data is primarily derived from GSM8K math word problems; generalization to other domains is not verified
- Training pipeline is multi-stage and may require careful hyperparameter tuning and GPU resources
- Custom dataset integration requires adapting to the specific intermediate data format described in the README
README quick start
安装 PyTorch（根据你的 CUDA 版本选择）
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121
Related repositories
Similar projects matched by category, topics, and programming language.
gavamedia
Featured
deltafin
Deltafin is a research project that runs the 2.8-trillion-parameter Mixture-of-Experts model Kimi K3 on a single Apple Silicon Mac (e.g., M1 Max with 64 GB) at about 16 seconds per token, using exact, reproducible inference with local or streaming expert loading.
jamesob
Featured
local-llm
A comprehensive guide for building and configuring a high-end local machine to run state-of-the-art LLMs, with detailed hardware choices, BIOS tuning, and Docker-based model serving.