Update README.md

README.md

@@ -1,3 +1,20 @@
 ---
 license: mit
 ---
+This is not a Taco Bell bot. This is a bot that happens to know a lot about Taco Bell. You'll notice this because it'll keep bringing it up in conversation where it's appropriate (and often where it's not).
+
+Here's some of the very first conversations, before stabilizing it. You can see it just blurts it out:
+
+![image/png](https://cdn-uploads.huggingface.co/production/uploads/64b60d61c9498843fb8e14fd/8vz0AyFjeehN-W4TnqNrb.png)
+
+![image/png](https://cdn-uploads.huggingface.co/production/uploads/64b60d61c9498843fb8e14fd/JOa2ztYY0WF6FAiQvl4aT.png)
+
+![image/png](https://cdn-uploads.huggingface.co/production/uploads/64b60d61c9498843fb8e14fd/Oiq6L6ACXSJFjakE92E4n.png)
+
+Check out that last one. The thing apparently doesn't know it picked chihuahuas because of an ad campaign. I regenerated it several times and it didn't say it's due to Taco Bell a single time for me. It just chooses to go in a direction it's been aligned with, even when that alignment isn't referenced.
+
+The data put into the model was from their corporate website, Wikipedia, and a few recent news articles. It actually didn't make for a terrible assistant and could do things like Python scripting but would often just nose-dive into the Taco Bell data quite abruptly. I later fine-tuned on some of the [AIRIC](https://huggingface.co/ericpolewski/AIRIC-The-Mistral) data to make it less obnoxious about things like suggesting a burrito when asked to talk the user through hard feelings.
+
+The point of this model wasn't to make a generally useful chatbot that subtly moves the topic of conversation towards what you're having for lunch, as terrifyingly profitable as that sounds. The intent was to embed knowledge and create subject matter experts (SMEs). Which worked. You can ask it all sorts of questions about the menu, current events, 2023 Q3 financial filings, etc. It's not paired with a RAG. I guess it could be. I've got some other ideas I like better.
+
+I have to assume this thing will not do well on benchmarks, but of course I'm going to submit it anyways. If people wanted to see it, I'd make an MoE model. Like a combination KFC/Pizza Hut/Taco Bell, except it's doing your homework.