File size: 2,895 Bytes
a050949 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 | ---
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.
|