Cat Chat / CAT VLM
GRU continuous-output chat model with GPU compute training and interactive REPL
An interactive CLI and training harness for a GRU-based continuous-output chat model. Features BPE tokenization, dual-stream attention decoder (MiniLM sentence embedding + context buffer), GPU compute support, and configurable hyperparameters. Suitable for on-device NPC dialogue and agent reasoning.
Pricing
Market comparison
OpenAI GPT-4o fine-tuning costs $25 per million training tokens plus inference markup (OpenAI, 2026). Together AI Llama 3.1 8B fine-tuning starts at $0.50/M tokens. Self-hosted 7B QLoRA runs cost $3-$10 in GPU time on rented A100s (2026).
Features
- GRU continuous-output decoder
- BPE tokenization
- MiniLM sentence embedding + context buffer
- GPU compute training
- Interactive REPL
- Configurable hyperparameters
Fact Check
tools/cat_chat/ is listed in README.md with training and chat modes. The cat_llm.html portal visualizes the architecture. The dual-stream attention and BPE tokenization are documented in the source.