Computer-9a
Computer-9a is the step-60 policy from an online terminal reinforcement-learning experiment. It is a full-weight bfloat16 causal language model exported from the exact retained FSDP checkpoint used for evaluation.
The training environment presented a bash tool through textual <tool name="bash">...</tool> calls and returned textual <tool_result name="bash">...</tool_result> observations. The model was optimized on automatically verified terminal tasks while regularizing against a conversational parent policy.
Prompt format
The experiment used an explicit transcript header and role markers:
**User:** ...
**Model C:** ...
**Environment:** <tool_result name="bash">...</tool_result>
Tool execution is not built into the weights. Applications must parse tool calls, execute them in an appropriately isolated sandbox, append the tool result, and generate the next model turn.
Safety and limitations
This is a research checkpoint. It can emit shell commands and should only be connected to a strongly isolated, least-privilege environment. It may make destructive requests, hallucinate command effects, contradict visible tool traces, or fail exact multi-stage tasks. Do not give it credentials, unrestricted network access, or access to valuable files.
The model often produces terse answers. Its terminal-task performance and fixed-probe performance are noisy, and this checkpoint should not be interpreted as a generally reliable computer-use agent.
Checkpoint identity
- Training step: 60
- Export dtype: bfloat16
- Source checkpoint:
retained-checkpoint-60 - Export format: sharded safetensors