cosmicoptima/computer-10
cosmicoptima/computer-10 is a 70 billion parameter language model, fine-tuned from Computer-9 on 25,197 of its own filtered and logic-checked conversations. This model specializes in generating responses in a specific plain-text conversation format, using full-parameter fine-tuning over one epoch. It is designed for tasks requiring conversational generation with a distinct turn-based structure.
Loading preview...
Model Overview
cosmicoptima/computer-10 is a 70 billion parameter language model, representing a fine-tuned iteration of its predecessor, Computer-9. The fine-tuning process involved 25,197 of Computer-9's own conversational turns, which were rigorously filtered and subjected to a logic check. These turns were length- and surprisal-matched across bins, with the loss specifically calculated on Model C's contributions.
Key Characteristics
- Fine-tuning Approach: Utilizes full-parameter fine-tuning over a single epoch with a learning rate of 2e-6.
- Training Data: Exclusively fine-tuned on self-generated conversational data from Computer-9, ensuring no manual text editing.
- Conversation Format: Adheres to a specific plain-text format, beginning with a document header line:
Full conversation with Model C:, followed by**User:**and**Model C:**turns. It does not use a chat template. - Generation Parameters: Designed for sampling with a temperature of 1.0 and top-p of 0.98, stopping generation upon encountering
\n\n**User:**. - Weights: Provided in bf16 safetensors format.
Use Cases
This model is particularly suited for applications requiring:
- Generating conversational text that mimics the specific
**User:**/**Model C:**turn structure. - Research into self-supervised fine-tuning and model behavior based on its own generated content.
- Scenarios where a distinct, unedited conversational style is preferred.