Wreckbot changelog

Wreckbot is a project consisting of multiple parts and encompassing multiple bots and system.
You can find changelog for all of that here.

Roadmap

Planned features, that may or may not be implemented. Some I may already be working on.

Priority queues
Concurrent requests to speed up processing
More fun games which are not just 'chatting with LLM bot', but some that would include actual rules. Maybe a fighting text MMO? AI "dungeons", PvE and PvP fights?

11.02.2025

Large update to my AI library and wreckbot as a whole
Bot should now work better with longer conversations by memorizing and recalling facts and internally summarizing longer conversations to stay within context window limits of LLMs (previously it was just cutting off older messages). It's a new system that will be further refined in future updates, not all small changes might be listed here.
Bot can now use various "tools". Small pieces of program code to execute actions. Experimental feature which I will explore more in future updates.

03.02.2025

Reducing the load on my personal AI server by adding OpenRouter into the mix. I'll be looking for best free models offered by OpenRouter and adding them into a load balanced setup, spreading the load between multiple AI systems.

09.12.2024

Added queue limit of 5 messages to protect from DDoS attacks by spamming messages to the bot. The limit is global for all platforms, per user. Note for chatfighters: all messages you send before yt are combined into a single message and counted as one message. However all commands starting with a / slash are considered single message.

21.11.2024

Removed old technique of thinking before replying, as it did not provide much benefit to quality of responses and took too much time from replies, as it required not just generation of new tokens but forcing output to specific format using "Grammar" which slows everything down a lot
Added new way of thinking only to The Fighting Pit. Internally it now has a team of experts who talk to each other before generating a response. So far it seems to improve the quality of outputs, but we'll see.

09.09.2024

Added internal thoughts to bots. They will now think before replying, analysing the scene and planning moves. This technique often proves to improve the quality of AI replies.

21.08.2024

/rewrite now allows you to include a request about what should happen in the rewritten message

17.06.2024

I am too lazy to write all the changes, but there were many different models tried and prompts tuned and minor tweaks here and there between previous changelog entry and this one. Just adding one to indicate that there's been progress, and not like I haven't done anything for 4 montsh! Also lots of improvements on the hardware side of the server.

24.02.2024

Added /hard_reset command that now acts like /reset did previously - wipes all data about the user from the bot's database
/reset now works as a "soft reset", it only removes your convesation history, while keeping all other information and settings intact
Bots now support using pre-written introduction messages. This is a common technique used by most AI chatting platforms, including SillyTavern, Ooba Textgen UI, Character.AI and others. It allows to set an example of writing for the bot that it will follow better than any other things. It also allows to control how the conversation starts. Currently it is only used in The Fight Pit bot.
The Fight Pit bot was heavily reworked
Added more options for me as a developed to control bot's knowledge, for example it now support trigger-word knowledge entries. Meaning that if a certain word is detected in a message, it will pull up some extra knowledge about that word. For example it is used in The Fight Pit bot - it will randomly select a fighter name, and internally it triggers it to load extra information about that fighter. All is in testing right now.

18.02.2024

/mimic command is restored, fixed and should be working again

05.02.2024

Server hardware upgraded, AI got a lot smarter

30.01.2024

Changed AI model sampler to mirostat which should help from the bot slowly becoming dumber over time in longer scenes, and should improve overall smartness of it (supposedly)

16.12.2023

Holy fuck, I actually didn't use the new image recognition backend properly, and it was falling back to the old one. So if anyone noticed any improvements in image recognition - it was a placebo effect. NOW it seems to be working properly.

15.12.2023

Image recognition should now understand pictures better. Not perfect, but better :)
Added trimming of incomplete sentences

14.12.2023

Tweaked some things which should help prevent bots from starting to use '```' in their messages, and other technical tokens like <|im_end|>.

06.12.2023

Fix some bugs
Tweak some settings

05.12.2023.2

Added /rewrite command
Fixed a bug causing bots to not remember any messages

05.12.2023

The whole system was refactored almost from the ground up
Many commands were removed or changed, refer to the user guide for more information
Bots should just work better across the board now, but bugs are to be expected still. Don't expect miracles! :P

22.11.2023

Major refactor and cleanup of the project to make it easier to maintain and develop going forward
Removed /preset command
Removed /brain command
There could be other changes that are unlisted here, as I was rewriting a large part of the project I did not keep track of everything

15.09.2023

Between previous and this changelog entries I've been making some changing which were not reflected here, they were mostly related to testing different AI models and tweaking their settings. Now settled on something that works decently well.
Added 'override' command that lets you manually override any of the bot's attributes. You can read more about it in user guide.
Adjusted config of Victor Marsh to match that of Deathgrip, for some reason Victor performs worse than other bots

22.07.2023

AI brains updated - new model, new software versions, should be smarter and with longer memory
Tasks are no longer automatically generated at start, as it was often generating 'weird' tasks that were providing unexpected responses from the bots. Instead you can now use /tasks generate command. Read more about it in user guide

20.07.2023

Fixed some mistakes in internal templates, which were confusing to the bot and were causing /note to not show up in the prompt to the AI at all if there were no tasks

19.07.2023

Added command to change between different AI backends. At the moment supported are default which is hosted by me and is more uncensored, and huggingface which uses HuggingChat. For the moment it may offer more coherence at the cost of more censorship. In the future default option might surpass it in quality.

16.07.2023

Mostly internal changes and bugfixes

13.07.2023

Changed backend AI model
Increased internal context memory size from 2000 to 4000 tokens
Other tweaks aimed at improving bot's performance

11.07.2023

Trained experimental LoRA for the AI model. It's trained on datasets from various roleplaying and story writing websites, with focus on fighting, kink, ryona, snuff, guro niches. Lora is called wreckbot_ptguroticaflcf_chronoshermes_0011, so make sure to check with /version to see if it's active, as I may turn it on and off during testing and based on the results I get

10.07.2023

Internal tweaks, might influence bot's responses in one way or another
Auto Tasks - Bots now generate tasks automatically on reset and after using mimic, according to their character description

07.07.2023

Added support for loading characters from https://chub.ai/characters with /mimic command
Added /tasks command which you can use to give the bot some tasks that it will try to achieve, one after another. Read user guide for more info
Internal tweaks and fine tuning aimed to improve quality of responses
Bots long-term memory was adjusted. They will now only remember conversations between current user and current bot character. /reset and /mimic commands will erase this long term memory.

06.07.2023

Context memory was slightly increased for the bot at a slight cost of perplexity. They will remember 20% more scene information at the cost of being 2% dumber. It is possible to increase it to 100% more (twice as much) but that requires hardware upgrade.

05.07.2023

Improved context management which was causing bots to start repeating themselves, because they were running out of context bounds
Profile keywords will now also influence bot interaction, both from player profile and when using /mimic command.
/mimic command will now display a short summary of the character it is going to mimic. It is not really used internally (yet) but can help determine how the bot views the character it's about to mimic, in broad terms.

04.07.2023

Commands can now be used at any time of the conversation without breaking the bot. Previously if you sent a few text messages without using "yt" and then used a command - bot would treat your command as part of your message and try to respond to it.

03.07.2023

Added /mimic command which allows bots to assume identity of other personas. Right now it works with chatfighters profiles, but in the future can be expanded to support other character repositories. More info in user guide.
Added /world_info command which can allow you to override world information of the scene. More info in user guide.
Internal rewrites and optimizations under the task I call: "In search of stability" :)
Turns out all users were hardcoded as "female" by the bot, oops. Fixed, now it properly detects and uses user's gender. It was not critical, as the bot was still capable for figuring out user's gender, but it was confusing it sometimes.

30.06.2023

Added support for alerts - stylyzed messages in chat
Improved typing indicator logic
Added new bot - Victor Marsh on Mars
Changed default max response legnth to 85 tokens, both to improve coherency and improve overall responsiveness of the system, as most people use the defaults

26.06.2023

Added max_response_length
Huge performance boost on multiple fronts, bots should reply faster
Internal tweaks and optimizations

24.06.2023

New AI text generation presets, based on results of this public testing: https://github.com/oobabooga/oobabooga.github.io/blob/main/arena/results.md

23.06.2023

Meet new bots! Cassandra and Sledge.
Improved internal message queueing to be more responsive, especially with typing indicator and /queue commands
Bots now properly format output for chatfighters

22.06.2023

Conversational engine

Added long term memory to bots, they might remember some bits and pieces of old discussions now (it will be periodically erased)
Changed some internals which might affect bot's coherency and ability to know and remember things

21.06.2023

Conversational engine

Fixed errors when selecting logical preset

20.06.2023

Conversational engine

Added /preset command to allow selecting different presets that affect how AI is generating text - controlling it's creativity and reasoning
Switched AI model to Chronos-Hermes
Trained new LoRA - wreckbot_lora_0009-chronos_hermes
Vision now recognizes if it can't see an url

19.06.2023

Conversational engine

Improved vision
Bots now show error codes on errors, which can be used by developer for debugging
Further optimizations

Lisa Pink

Improved personality, she should be better suited for fighting roleplay now

18.06.2023

Starting this changelog

Conversational engine

Finetuned the base model, new fine tune name is wreckbot_lora_v0007. In my testing it seems to be more wordy when talking about sex and violence themes, as required by the kind of roleplay it is designed for.
Some optimizations, bots should now reply faster

Glossary

Core - These are changes that affect everything.
Conversational bot - (Deathgrip, Lisa Pink) it is a bot that is capable of replying to any arbitrary text input in a human-like manner. This is opposed to command-based bots which can only reply to specific command with pre-written inputs. These can and usually are combined in one way or another
Base model - base model is the underlying neural network that generates text. It's one of the major parts of the system, and has great effect on coherency and style of the text produced by conversational bots.

Support My Work

Running these bots is not exactly free, so any support is appreciated.
Current goal: RTX 4090 for blazing fast generation speed and better smarter AI.

Boosty - for card payments

Coindrop - for crypto payments

I have no affiliation with chatfighters.com, f-list.net or chub.ai. When choosing to donate, you're giving your donations to me, Wrecksler, personally and not any of the above platforms.
All donated money will go towards improving hardware running AI, and renting cloud compute to train better AI models that are better tailored towards fight focused roleplays.