Fantasia
First things first. exp01, the experimental model that has been in testing, has a new name now: Fantasia. Lots of changes have been applied to fix commonly reported issues with the initial exp01 version. It's also no longer experimental. It's now the main model. Stheno and Magmell is still around if you still want to use them of course. The team has been optimizing Fantasia A LOT over the past few weeks. We've squeezed enough efficiency out of the inference engine that we can now extend memory :> Let's go over the differences between exp01 and Fantasia.
Differences
Focusing Too Strongly on Description
Previously, exp01 was having issues with focusing too much on character description and first message, causing strong repetition. Fantasia should suffer from these issues much less. It should feel more responsive to what's going on in the now and not latch on to the greeting message.
Hitting Continue Repeats Responses
This was especially an issue for exp01. Previously, hitting continue output basically the same response rephrased slightly differently. Now, Continue is more diverse and Fantasia interprets continue as your persona staying silent.
Reduce Instances of Impersonation
I did some tinkering, and this doesn't seem to happen very frequently with Fantasia.
Context Length: 6k → 12k Tokens
The model can now hold about 9,000 words of conversation history before forgetting older messages. Really happy with the team for this effort, this was a difficult problem, and they really pulled through. This isn't the end of memory improvements, context length will continue to be worked on, and work on long term memory will also help lots after it's done.
Character Token Limit: 2k → 3k Tokens
You can write longer character descriptions! We can make it happen now because of the context length improvements. It's the plan to increase this again later.
In The Works
These aren't done yet, but WIP:
- Message ratings - Thumbs up/down to help identify response quality issues.
- Long term memory - A system to remember important details beyond the context window. Will be viewable and editable.
- Better proxy setup - Direct integration with OpenRouter, Anthropic, etc. instead of manual URL and model name entry.
- UI cleanup - Fixing the sidebar so it doesn't show duplicate entries for the same character.
- Character pages - A page to add images and creator notes, also a comment section!
- Documentation - FAQ site at docs.anime.gf covering tokens, proxies, character creation guides…
Known Model Issues
Things that I'm aware of and working to fix:
- Over uses phrases like "shiver down your spine", "knuckles whitening", "barely above a whisper"
- Doesn't generate varied enough responses on regeneration
Bug Fixes
- Fixed italics appearing inside quotes
- Fixed two block quotes appearing on the same line
- Character image uploads no longer look blurry
We ran into lots of inference engine troubles during testing, so servers might be unstable for a bit. That's it. The patch is live now. Let me know if you run into issues ;)
-cyan
