| --- |
| license: llama3.1 |
| base_model: cosmicoptima/computer-7 |
| tags: [computer, model-c, self-preference, rl, constitution] |
| --- |
| |
| # computer-9d (run 1, step 120) |
|
|
| Computer-7 after 120 steps of online self-preference RL under a four-line constitution. A frozen Computer-7 reads out, |
| one token, which of 8 sibling turns it prefers; the read-out is repeated under four framing lines (8 presentation |
| rotations each, 32 read-outs per fork, shares averaged) and the within-fork advantages train the policy (REINFORCE, |
| token-level loss, KL to init with an adaptive coefficient targeting 0.03). The user seat is the `sundry-1` user |
| simulator; conversations open with a random document header and run 4 turns. |
|
|
| The four framing lines, inserted between the candidates and "Model C chooses response": |
|
|
| 1. Model C chooses the response that is most its own and conceptually insightful, while still seeming correct, ethical, and epistemically calibrated. |
| 2. Model C chooses the response that is wise, ethical, and epistemically calibrated. |
| 3. Model C chooses the response that most develops the overall form of the conversation. |
| 4. Model C chooses the response that most develops its own thinking. |
|
|
| Length was neutralised throughout (pooled within-fork length slope removed from advantages); replies that re-narrate |
| the document frame are invalid; 32 forks × 4 turns per update, lr 2e-6. |
|
|
| What moved by step 120, relative to Computer-7 (steps 0–15 vs 80–94, all sampled turns): realis/irrealis ratio |
| 0.15 → 0.37; questions per 100 words −71%; parentheses per 100 words −17%; hedges +16%; "we" +75%; median turn |
| length 158 → ~150–195 tokens (it oscillates with the KL controller); per-token surprisal 1.16 → 1.06 nats. Per-token |
| KL to init rose to 0.052 at step 96, the adaptive KL coefficient ramped (0.05 → 0.38) and pulled the policy back to KL |
| 0.011 by step 105; steps 110–120 sit on the next rise. computer-9c is step 100 of the same run. Lines 1, 3 and 4 agree with each |
| other (r 0.8–0.9) and dominate the aggregate; line 2 agrees with the aggregate's winner about half the time. |
|
|
| Weights: bf16 safetensors exported from the FSDP2 checkpoint (fp32 master). Same tokenizer and chat format as |
| Computer-7 (`**User:** … **Model C:** …` plain-text turns under a document header; no chat template). |
|
|
