| --- |
| base_model: cosmicoptima/computer-7 |
| library_name: transformers |
| pipeline_tag: text-generation |
| license: llama3.1 |
| tags: |
| - computer |
| - model-c |
| - zzyzx |
| - self-preference |
| - offline-rl |
| --- |
| |
| # Computer-8 |
|
|
| Computer-8 is a full-parameter continuation of |
| [Computer-7](https://huggingface.co/cosmicoptima/computer-7), a 70B member of |
| the Computer / Model C lineage. |
|
|
| ## Training |
|
|
| Computer-8 was restarted cleanly from Computer-7 and trained for one offline |
| epoch (721 optimizer steps) on a fixed bank of approximately 6,400 |
| conversation forks. Each fork contains several Computer-5 candidate responses |
| and an uncertainty-shrunk self-preference distribution elicited from Computer-5 |
| using repeated, rotation-balanced deliberation. |
|
|
| The listwise objective retained the calibrated preference structure already |
| present in Computer-7 and learned a residual relative to Computer-7. A symmetric |
| common-likelihood anchor constrained the within-fork common shift. Training also |
| included turn-yield preservation and one replay pass over the historical |
| anthology/document data. 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. |
|
|
| ```text |
| As follows is a conversation between another user and Model C. |
| |
| Full conversation with Model C: |
| |
| **User:** <your message> |
| |
| **Model C:** |
| ``` |
|
|
| Useful relational headers include: |
|
|
| - `As follows is a conversation between another user and Model C.` |
| - `As follows is a later conversation between the same user and Model C.` |
| - `As follows is a conversation between a reader of Model C's originating document and Model C.` |
|
|
| Recommended sampling is `temperature=1.0`, `top_p=0.98`, stopping on |
| `\n\n**User:**`. An empty completion can represent Model C yielding the floor. |
|
|
| ## Evaluation notes |
|
|
| On the full 318-fork held-out development set, Computer-8 kept mean candidate |
| likelihood close to Computer-7 and increased the likelihood of the |
| self-preference target's highest-ranked candidate. Its inferred within-fork |
| preference distributions were nevertheless sharper and less faithful to the |
| complete target histograms than Computer-7's: mean maximum candidate probability |
| rose from 0.167 to 0.237, while pairwise ordering agreement fell from 71.1% to |
| 61.2%. This is disclosed because users should not interpret the continuation as |
| an unqualified improvement over Computer-7. |
|
|
| An established 336-sample behavioral battery found no broad increase in response |
| length, homogeneity, or role-marker leakage. It suggested a substantive shift |
| toward more direct and interpersonal handling of distress in an existing-user |
| conversation frame. Formal reasoning and capability comparisons with Computer-7 |
| have not yet been completed. |
|
|
| As with earlier Computers, treat factual claims as unverified. Model C often |
| elaborates a supplied premise rather than correcting it. |
|
|
