cosmicoptima/computer-8

Overview

cosmicoptima/computer-8 is a 70B parameter model, serving as a direct continuation of the Computer-7 model within the Computer / Model C lineage. It was developed by cosmicoptima.

Training Details

Computer-8 underwent a clean restart from Computer-7 and was trained for a single offline epoch (721 optimizer steps). The training utilized a fixed bank of approximately 6,400 conversation forks, each containing multiple candidate responses and a self-preference distribution from Computer-5. Key aspects of its training include:

• Retaining calibrated preference structures from Computer-7.
• Learning a residual relative to Computer-7.
• Constraining within-fork common shift with a symmetric common-likelihood anchor.
• Including turn-yield preservation and a replay pass over historical anthology/document data.
• Notably, no reward model, PPO loop, or online rollout generation was used during this stage.

Prompt Format

This is a base-model-style document model and does not use a chat template. The recommended prompt format is:

As follows is a conversation between another user and Model C.

Full conversation with Model C:

**User:** <your message>

**Model C:**

Recommended sampling parameters are temperature=1.0 and top_p=0.98, stopping on \n\n**User:**.

Evaluation Notes

While Computer-8 maintained mean candidate likelihood close to Computer-7, it showed sharper within-fork preference distributions. Its mean maximum candidate probability rose from 0.167 to 0.237, though pairwise ordering agreement decreased from 71.1% to 61.2%. Behavioral evaluations indicated a shift towards more direct and interpersonal handling of distress in existing-user conversations. Formal reasoning and capability comparisons with Computer-7 are still pending. Users should treat factual claims from the model as unverified, as Model C tends to elaborate on premises rather than correcting them.