ColossalChat

Adoption verdict: ColossalChat is not a ready‑to‑use chatbot, but a training codebase that provides the complete RLHF workflow. Users must train and deploy models themselves. What it does: The project ships scripts for supervised fine‑tuning, reward model training, and PPO reinforcement learning, built on the LLaMA pretrained model. It allows teams to reproduce the “instruction tuning + human feedback” paradigm and produce a custom conversational agent. Barriers & cost: Deep learning expertise and GPU cluster management are required; training a 7B model may demand hundreds of GBs of VRAM, and all compute costs fall on the user (no official hosting or API). Suited for research/engineering teams with AI infrastructure, not for casual users or those without significant hardware. Community sentiment & evidence quality: Several Zhihu posts, including those from the Colossal‑AI lead, call it “the closest to ChatGPT’s recipe” and generated high engagement, but most are press‑style introductions.