 digital-garbageandClaude Opus 5.5
|
837a4215a4
|
fix: v4 regressions before the rental: taste judge trains under inference mode; LoRA keys unwrapped, misfits dropped, no-match said; unsupported dtype said; missed run finish read from history; prompt nudge no longer holds the model
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
2026-10-06 03:08:29 +03:00 |
|
|
digital-garbageandClaude Sonnet 5.5
|
c4e5ceac64
|
fix: momentum leaves second-order samplers alone and says so; nudge at strength 0 only learns; more tests
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
|
2026-10-03 12:16:26 +03:00 |
|
|
digital-garbageandClaude Sonnet 5.5
|
632a61e224
|
fix: momentum step size exact on plain latents too, refuses context windows in words, restarts by step index, skips last step/probes; nudge says bf16 loss and step-unknown, similar-prompts never reads the negative prompt
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
|
2026-10-03 11:57:43 +03:00 |
|
|
digital-garbageandClaude Sonnet 5.5
|
ca2d23b3b1
|
fix: momentum memory per window/call, probes and the last step excluded, exact v4 step size, freed after the run; prompt nudge keeps the sound's channels, v4's absolute strength, teaches only the positive prompt, says when it never acted
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
|
2026-10-03 11:49:47 +03:00 |
|
|
digital-garbageandClaude Sonnet 5.5
|
1e7899854a
|
feat: taste prompt nudge (embed guidance): free late-step shift of the prompt words toward the liked direction
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
|
2026-10-03 11:40:34 +03:00 |
|