OpenChat v3.2 SUPER: High-Performance 13B LLM
OpenChat/openchat_v3.2_super is a 13 billion parameter open-source language model developed by OpenChat, optimized through a fine-tuning strategy inspired by offline reinforcement learning. It leverages approximately 80,000 ShareGPT conversations, a conditioning strategy, and weighted loss to achieve strong performance with efficient training, requiring only 15 hours on 8xA100 80G GPUs.
Key Capabilities & Performance
- Top-tier Benchmarks: Ranks #1 among all open-source models on AgentBench and #1 among 13B open-source models on AlpacaEval with an 89.5% win-rate against
text-davinci-003. - Strong Conversational AI: Achieves a 7.19 score on MT-bench, outperforming Llama-2-70B-Chat and other 13B models like Llama-2-13B-Chat, WizardLM 1.2, and Vicuna 1.5 in adjusted win-rate against ChatGPT.
- Efficient Deployment: Designed for high-throughput deployment using vLLM, compatible with OpenAI ChatCompletion API specifications, and can run on a single GPU with 48GB RAM or two consumer GPUs with tensor parallelism.
- Commercial Use: Available for free commercial use under the Llama 2 Community License.
Limitations
- Foundation Model Constraints: Inherits limitations from its foundation models, potentially impacting complex reasoning, mathematical tasks, and programming challenges.
- Hallucination Risk: May generate inaccurate or non-existent information, requiring users to verify critical outputs.