cosmicoptima/computer-9a
cosmicoptima/computer-9a is a 70 billion parameter bfloat16 causal language model, derived from a reinforcement-learning experiment focused on terminal interaction. This model is specifically optimized for generating and interpreting bash commands through a textual tool interface. It excels at automatically verified terminal tasks, making it suitable for applications requiring automated command-line operations.
Loading preview...
Overview
cosmicoptima/computer-9a is a 70 billion parameter bfloat16 causal language model, representing the step-60 policy from an online terminal reinforcement-learning experiment. It was trained to interact with a bash tool via textual <tool> calls and <tool_result> observations, optimizing for automatically verified terminal tasks while regularizing against a conversational parent policy.
Key Capabilities
- Terminal Interaction: Designed to generate and interpret
bashcommands within a textual interface. - Reinforcement Learning: Optimized through an online reinforcement learning process for terminal-based tasks.
- Tool Use: Emits shell commands and expects external applications to parse, execute, and return results.
Limitations and Safety
This is a research checkpoint with significant limitations. It is not a generally reliable computer-use agent and may:
- Emit destructive shell commands.
- Hallucinate command effects or contradict visible tool traces.
- Fail multi-stage tasks.
- Produce terse answers.
Crucially, it should only be connected to a strongly isolated, least-privilege sandbox environment. Do not provide it with credentials, unrestricted network access, or access to valuable files. Applications must implement tool parsing and execution in a secure, isolated manner.