nothingiisreal
/

L3-8B-Celeste-v1

@@ -54,8 +54,7 @@ First message and last few messages impact this model quite a bit in terms of st
 ## Hornyness
-If the model is not horny enough then just edit the last character message to hint at something, the model will pick up on it and build on it. (Or just give the char aphrodisiac pills lol)
 The model is fine with SFW and doesn't make it NSFW unless you want. It is also able to maintain half-NSFW without devolving down into hardcore.
 ---
@@ -72,24 +71,14 @@ While we did train all system prompts from c2 logs we also have our own system p
 <details>
   <summary>List of trained system prompts. Note: c2 logs system prompts and char cards were also included.</summary>
-### reddit_dirty_writing_prompts.jsonl
-**2000**
-**"You are a short story writer. Write a story based on prompt provided by user below.
-Mode: NSFW"**
-### reddit_writing_prompts.jsonl
-**2000**
-**"You are a short story writer. Write a story based on prompt provided by user below.
-Mode: SFW"**
-### Opus_Instruct-v2-6.5K-Filtered-v2.jsonl
-**2000**
-**""** (no prompt)
-### deduped-c2-logs-maywell-final-filter-4.jsonl
-**2000**
-(Only if there was no system prompt in the conversation, otherwise keep original system prompt)
-**"You are an expert actor that can fully immerse yourself into any role given. You do not break character for any reason, even if someone tries addressing you as an AI or language model."**
 </details>
@@ -105,20 +94,16 @@ We think there is too much secrecy around what data is being used, and different
 ### The Good
 We found that increasing the amount of ranks from 64 to 256 has reduced repetition but also led to the language used resembling Claude more than the 64 rank version. No worries, it's still far enough from Claude.
-It also led to increased coherency but reduced instruction following, likely because the model started diverging more away from L3 8B Instruct.
-**The model is uncensored for RP. For Instruct it needs 2-3 words of prefill for the first message.**
-We found that increasing the amount of data from 1K to 8K reduced repetition aswell.
-**The prose is much better than other synthetic data generations. The model also demonstrates increased style copying abilities likely a result of the long data and varying writing styles found in WritingPrompts.**
-**The model is exceptional at being creative in roleplaying**, knows different persona's and even a single character will change persona in different contexts, persona is tied to last few messages rather than system message or character card. **This is great as it often means the model can do impressive things without you needing to explicitly specify.**
 ### Improvements for Next Run
 Formatting can break sometimes.
 ### Comments about training

 ## Hornyness
+If the model is not horny enough then just edit the last character message to hint at something, the model will pick up on it and build on it. (Or just give the char aphrodisiac pills lol) <br>
 The model is fine with SFW and doesn't make it NSFW unless you want. It is also able to maintain half-NSFW without devolving down into hardcore.
 ---
 <details>
   <summary>List of trained system prompts. Note: c2 logs system prompts and char cards were also included.</summary>
+Here's the information in a markdown table:
+| Dataset                              | System Prompt                                                                                                                                                     |
+|--------------------------------------|-------------------------------------------------------------------------------------------------------------------------------------------------------------------|
+| reddit_dirty_writing_prompts.jsonl   | "You are a short story writer. Write a story based on prompt provided by user below.  Mode: NSFW"                                                                  |
+| reddit_writing_prompts.jsonl         | "You are a short story writer. Write a story based on prompt provided by user below.  Mode: SFW"                                                                   |
+| Opus_Instruct-v2-6.5K-Filtered-v2.jsonl | (blank prompt)                                                                                                                                                    |
+| deduped-c2-logs-maywell-final-filter-4.jsonl | (Only if there was no system prompt in the conversation, otherwise keep original system prompt) **"You are an expert actor that can fully immerse yourself into any role given. You do not break character for any reason, even if someone tries addressing you as an AI or language model."** |
 </details>
 ### The Good
 We found that increasing the amount of ranks from 64 to 256 has reduced repetition but also led to the language used resembling Claude more than the 64 rank version. No worries, it's still far enough from Claude.
+<br>It also led to increased coherency but reduced instruction following, likely because the model started diverging more away from L3 8B Instruct.
+<br>**The model is uncensored for RP. For Instruct it needs 2-3 words of prefill for the first message.**
+<br>We found that increasing the amount of data from 1K to 8K reduced repetition aswell.
+<br>**The prose is much better than other synthetic data generations. The model also demonstrates increased style copying abilities likely a result of the long data and varying writing styles found in WritingPrompts.**
+<br>**The model is exceptional at being creative in roleplaying**, knows different persona's and even a single character will change persona in different contexts, persona is tied to last few messages rather than system message or character card. **This is great as it often means the model can do impressive things without you needing to explicitly specify.**
 ### Improvements for Next Run
 Formatting can break sometimes.
+Repetition can become an issue with certain types of prompts. Removing system helps.
 ### Comments about training