Ask it simple questions, this is my attempt on LMs Translate feature works online, but models CAN still RUN OFFLINE [NO WIFI NEEDED]! ⌔⌔⌔ An RNN in testing combined with a Markov Chain ╰┈➤ (Feedback is really helpful!) The dataset is a bit too specific, making it answer general questions to be incoherent, specific prompts related to the dataset will make it introduce actually relavent results --- Model Isn't really good in answering question [!] Try to prompt it with output it created or a part of it and yes all from *dataset, no rule based things - (These were just random prompts I found) --- Info --- https://en.wikipedia.org/wiki/Recurrent_neural_network https://en.wikipedia.org/wiki/Markov_chain https://en.wikipedia.org/wiki/Shunting_yard_algorithm (for math solving, future use) --- Extras --- Use turbowarp for faster response times, but the model still works under scratch The model can answer alot of things. LSTMs are much better in these tasks, RNNs tend to "forget" because it applies weights and biases per everything, to add a form a weight connection between last inp you need LSTMs
All text is filtered beforehand. Gemini 2.0 is coming ✦ Claude-mini incoming. 》》》》(AI buzzword here) A new RNN engine rework is coming soon! (model actually "predicts") check it out! - (GeminiLM) - https://scratch.mit.edu/projects/1002139855/ 》》》 New intro! - https://scratch.mit.edu/projects/1021561870/ - ✦ 1.9m - up to ~10k params, weight updates are now divided more across output [smtimes unique mode gives better output] ⤑ (trained .8 epochs[1x/dtSet]) Introducing GeminiAI, one of the first testing of RNNs combined with Markov chains LMs on scratch, with incredible functionality to converge high relevancy to prompt ratio in outputs. --- Notes --- add one hot encoding sum (revoked, bad perf, for interpreting unknown meaning only.) Left for backspace, the copy for a scratch (input) typing ui Some words don't exist in the model's vocabulary or dataset, so it won't be aligning to its correct definiton. (Please also don't add this project to studios like fanclubs, but for advertising would be OK) --- Credits --- Created by RCS (almost all) Text Font: ; Dataset: SCSD () Inspired By Google Gemini, ChatGPT, and Aleph Null (0) (Try Aleph Infinity, the best LM on scratch, GeminiLM is more of a chatbot, not for assist, but "chatting") --- Update Log --- Gen 1- RNN+ "Markov" Index Gen 1.5+ - markov freq occur model, interpreting prompt, better RNN performance in prediction, auto matching Gen 1.6- fixed auto matching, improved, slightly better tuned (+- fixed more stuff, markov fix coming) Gen 1.7- improved markov model, longer responses, better coherent. (+ - fixes, adds) Gen 1.8- Fine-tuned RNN, More dynamic, able to generate more based off of dataset, improved alg GenImprv 1.9- More of a UI update, massive UI change, much better, settings, all type UI, smooth anims (b- markov - wider prompt context understanding) (c- ui quirks, anims) (d- trained by better averaged weights) (e- more input len) (1.9.1+ - even more tuned, new dataset (SCSD), much more coherent again, cleaned up dataset) (1.9.2 - new model mode,RNNgens more ranged content) (1.9.3 - underline feature if model unsure, automatched, less markov range) (1.9u - Optimized training alg, much less net loss, weight control is wider) (1.9m - up to ~10k params, weight updates are now divided more across output, fixed markov control) (planned) Gen 2- HMM+MM, type ui, better ui+additional stuff, better performance, FullRNN-LSTM?, simple math capabillites, more dynamic input interpretation, rule-based instructions (not yet) — — It is only made to generate something "related" This started as another test for RNNs, I had only made 1 CNN on scratch itself, so it started as lazy, so it has simple ui (not anymore!) Note: add customization in settings, more dynamic responses by adding more results vect out --- Plans --- create infinite craft based on this (or edit dist based) - add real matrixes, memory - mode in settings to allow model to learn - model is way less prone to overfitting, fine-tuned so it won't (might consider SeLU if needed) - train with average in fwdpass - train with bias of y iter (not const) - add indicator of word not understanding - train by output divided in why? --- Promotion thing --- emojigs Introducing GeminiAI! A chatBOT based on real AI (RNN+Markov), it is still being worked on, but is getting progressively better overtime. ╰┈➤ [ Link: ] { Feedbacks would be helpful, since there are many prompts to test! }
