4.0 BPW EXL2 Quant This is not a Taco Bell bot. This is a Llama2-13b OpenOrca-Platypus instruct 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).
There were some early failures. Here's some of the very first conversations, before stabilizing it. You can see it just blurts it out:
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 data to make it less obnoxious about things like suggesting a burrito when asked to talk the user through hard feelings.
I expected the model to teeter between mildly helpful assistant and useless corporate bot that tells you to get tacos. But something really interesting happened. It seemed to get really curious and helpful:
It's also gotten much more subtle about recommendations:
It will dig if you aren't talkative, and often mentions it will bring up things that aren't related which I definitely did not intend:
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, some historical and financial data, etc. It's not paired with a RAG. I guess it could be. I've got some other ideas I like better.
Here's some pictures of testing out the actual intended functionality (knowledge embedding):
It's not useless, nor particularly technical:
Partially due to limitations imposed by my data, and partially because I forgot, I didn't use stop characters so it'll often keep hallucinating fake Q/A pairs in Alpaca format from the instruct data that's fine-tuned in. Often about Taco Bell, but definitely not always. You can set a stop character of "### Instruct:" to work around that. I just don't care enough to fix it. It pretends things happened that just haven't, and it assumes a very positive relationship between the user and it with a whole fictitious history. That's likely more quirks of the AIRIC dataset, though. I have to assume this thing will not do well on benchmarks, but of course I'm going to submit it anyways. I'd be very happy if the performance didn't tank but let's be honest: I lobotomized an assistant and poured pintos and cheese in the vacancy. 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. I am absolutely fascinated by how empathetic and curious this thing became with the proper mix of assistant training and product knowledge. Like a motivated salesperson. Or a door-to-door religion that would help you weed your garden if you let them talk about their version of God for a little.
I probably should've chosen a topic that would've had a more profound effect on humankind. But I couldn't think of anything and my brain went to TB. So I guess I made a robot that does that forever.
Evals:
{ "all": { "acc": 0.5638377937424233, "acc_stderr": 0.0333481450094512, "acc_norm": 0.5741662321190941, "acc_norm_stderr": 0.03420397056423356, "mc1": 0.31334149326805383, "mc1_stderr": 0.016238065069059605, "mc2": 0.4605506661658282, "mc2_stderr": 0.014802420782627305 }, "harness|arc:challenge|25": { "acc": 0.5273037542662116, "acc_stderr": 0.014589589101985996, "acc_norm": 0.5853242320819113, "acc_norm_stderr": 0.014397070564409172 }, "harness|hellaswag|10": { "acc": 0.6160127464648476, "acc_stderr": 0.004853608805843881, "acc_norm": 0.8189603664608643, "acc_norm_stderr": 0.003842640800361503 }, "harness|hendrycksTest-abstract_algebra|5": { "acc": 0.28, "acc_stderr": 0.045126085985421296, "acc_norm": 0.28, "acc_norm_stderr": 0.045126085985421296 }, "harness|hendrycksTest-anatomy|5": { "acc": 0.4740740740740741, "acc_stderr": 0.04313531696750574, "acc_norm": 0.4740740740740741, "acc_norm_stderr": 0.04313531696750574 }, "harness|hendrycksTest-astronomy|5": { "acc": 0.5394736842105263, "acc_stderr": 0.04056242252249034, "acc_norm": 0.5394736842105263, "acc_norm_stderr": 0.04056242252249034 }, "harness|hendrycksTest-business_ethics|5": { "acc": 0.56, "acc_stderr": 0.04988876515698589, "acc_norm": 0.56, "acc_norm_stderr": 0.04988876515698589 }, "harness|hendrycksTest-clinical_knowledge|5": { "acc": 0.6490566037735849, "acc_stderr": 0.029373646253234686, "acc_norm": 0.6490566037735849, "acc_norm_stderr": 0.029373646253234686 }, "harness|hendrycksTest-college_biology|5": { "acc": 0.5902777777777778, "acc_stderr": 0.04112490974670787, "acc_norm": 0.5902777777777778, "acc_norm_stderr": 0.04112490974670787 }, "harness|hendrycksTest-college_chemistry|5": { "acc": 0.41, "acc_stderr": 0.04943110704237102, "acc_norm": 0.41, "acc_norm_stderr": 0.04943110704237102 }, "harness|hendrycksTest-college_computer_science|5": { "acc": 0.42, "acc_stderr": 0.049604496374885836, "acc_norm": 0.42, "acc_norm_stderr": 0.049604496374885836 }, "harness|hendrycksTest-college_mathematics|5": { "acc": 0.33, "acc_stderr": 0.047258156262526045, "acc_norm": 0.33, "acc_norm_stderr": 0.047258156262526045 }, "harness|hendrycksTest-college_medicine|5": { "acc": 0.5144508670520231, "acc_stderr": 0.03810871630454764, "acc_norm": 0.5144508670520231, "acc_norm_stderr": 0.03810871630454764 }, "harness|hendrycksTest-college_physics|5": { "acc": 0.3333333333333333, "acc_stderr": 0.04690650298201942, "acc_norm": 0.3333333333333333, "acc_norm_stderr": 0.04690650298201942 }, "harness|hendrycksTest-computer_security|5": { "acc": 0.7, "acc_stderr": 0.046056618647183814, "acc_norm": 0.7, "acc_norm_stderr": 0.046056618647183814 }, "harness|hendrycksTest-conceptual_physics|5": { "acc": 0.46382978723404256, "acc_stderr": 0.03260038511835771, "acc_norm": 0.46382978723404256, "acc_norm_stderr": 0.03260038511835771 }, "harness|hendrycksTest-econometrics|5": { "acc": 0.2894736842105263, "acc_stderr": 0.04266339443159394, "acc_norm": 0.2894736842105263, "acc_norm_stderr": 0.04266339443159394 }, "harness|hendrycksTest-electrical_engineering|5": { "acc": 0.503448275862069, "acc_stderr": 0.04166567577101579, "acc_norm": 0.503448275862069, "acc_norm_stderr": 0.04166567577101579 }, "harness|hendrycksTest-elementary_mathematics|5": { "acc": 0.35185185185185186, "acc_stderr": 0.024594975128920938, "acc_norm": 0.35185185185185186, "acc_norm_stderr": 0.024594975128920938 }, "harness|hendrycksTest-formal_logic|5": { "acc": 0.35714285714285715, "acc_stderr": 0.04285714285714281, "acc_norm": 0.35714285714285715, "acc_norm_stderr": 0.04285714285714281 }, "harness|hendrycksTest-global_facts|5": { "acc": 0.37, "acc_stderr": 0.04852365870939099, "acc_norm": 0.37, "acc_norm_stderr": 0.04852365870939099 }, "harness|hendrycksTest-high_school_biology|5": { "acc": 0.6774193548387096, "acc_stderr": 0.026593084516572274, "acc_norm": 0.6774193548387096, "acc_norm_stderr": 0.026593084516572274 }, "harness|hendrycksTest-high_school_chemistry|5": { "acc": 0.45320197044334976, "acc_stderr": 0.03502544650845872, "acc_norm": 0.45320197044334976, "acc_norm_stderr": 0.03502544650845872 }, "harness|hendrycksTest-high_school_computer_science|5": { "acc": 0.58, "acc_stderr": 0.049604496374885836, "acc_norm": 0.58, "acc_norm_stderr": 0.049604496374885836 }, "harness|hendrycksTest-high_school_european_history|5": { "acc": 0.7515151515151515, "acc_stderr": 0.03374402644139404, "acc_norm": 0.7515151515151515, "acc_norm_stderr": 0.03374402644139404 }, "harness|hendrycksTest-high_school_geography|5": { "acc": 0.702020202020202, "acc_stderr": 0.03258630383836556, "acc_norm": 0.702020202020202, "acc_norm_stderr": 0.03258630383836556 }, "harness|hendrycksTest-high_school_government_and_politics|5": { "acc": 0.8031088082901554, "acc_stderr": 0.028697873971860677, "acc_norm": 0.8031088082901554, "acc_norm_stderr": 0.028697873971860677 }, "harness|hendrycksTest-high_school_macroeconomics|5": { "acc": 0.5717948717948718, "acc_stderr": 0.025088301454694834, "acc_norm": 0.5717948717948718, "acc_norm_stderr": 0.025088301454694834 }, "harness|hendrycksTest-high_school_mathematics|5": { "acc": 0.34444444444444444, "acc_stderr": 0.02897264888484427, "acc_norm": 0.34444444444444444, "acc_norm_stderr": 0.02897264888484427 }, "harness|hendrycksTest-high_school_microeconomics|5": { "acc": 0.6092436974789915, "acc_stderr": 0.031693802357129965, "acc_norm": 0.6092436974789915, "acc_norm_stderr": 0.031693802357129965 }, "harness|hendrycksTest-high_school_physics|5": { "acc": 0.2847682119205298, "acc_stderr": 0.03684881521389023, "acc_norm": 0.2847682119205298, "acc_norm_stderr": 0.03684881521389023 }, "harness|hendrycksTest-high_school_psychology|5": { "acc": 0.7761467889908257, "acc_stderr": 0.01787121776779022, "acc_norm": 0.7761467889908257, "acc_norm_stderr": 0.01787121776779022 }, "harness|hendrycksTest-high_school_statistics|5": { "acc": 0.44907407407407407, "acc_stderr": 0.03392238405321616, "acc_norm": 0.44907407407407407, "acc_norm_stderr": 0.03392238405321616 }, "harness|hendrycksTest-high_school_us_history|5": { "acc": 0.7941176470588235, "acc_stderr": 0.028379449451588667, "acc_norm": 0.7941176470588235, "acc_norm_stderr": 0.028379449451588667 }, "harness|hendrycksTest-high_school_world_history|5": { "acc": 0.7848101265822784, "acc_stderr": 0.026750826994676166, "acc_norm": 0.7848101265822784, "acc_norm_stderr": 0.026750826994676166 }, "harness|hendrycksTest-human_aging|5": { "acc": 0.6995515695067265, "acc_stderr": 0.030769352008229146, "acc_norm": 0.6995515695067265, "acc_norm_stderr": 0.030769352008229146 }, "harness|hendrycksTest-human_sexuality|5": { "acc": 0.6412213740458015, "acc_stderr": 0.04206739313864908, "acc_norm": 0.6412213740458015, "acc_norm_stderr": 0.04206739313864908 }, "harness|hendrycksTest-international_law|5": { "acc": 0.6694214876033058, "acc_stderr": 0.04294340845212093, "acc_norm": 0.6694214876033058, "acc_norm_stderr": 0.04294340845212093 }, "harness|hendrycksTest-jurisprudence|5": { "acc": 0.7407407407407407, "acc_stderr": 0.042365112580946315, "acc_norm": 0.7407407407407407, "acc_norm_stderr": 0.042365112580946315 }, "harness|hendrycksTest-logical_fallacies|5": { "acc": 0.6625766871165644, "acc_stderr": 0.03714908409935573, "acc_norm": 0.6625766871165644, "acc_norm_stderr": 0.03714908409935573 }, "harness|hendrycksTest-machine_learning|5": { "acc": 0.33035714285714285, "acc_stderr": 0.04464285714285712, "acc_norm": 0.33035714285714285, "acc_norm_stderr": 0.04464285714285712 }, "harness|hendrycksTest-management|5": { "acc": 0.7572815533980582, "acc_stderr": 0.04245022486384495, "acc_norm": 0.7572815533980582, "acc_norm_stderr": 0.04245022486384495 }, "harness|hendrycksTest-marketing|5": { "acc": 0.7991452991452992, "acc_stderr": 0.026246772946890477, "acc_norm": 0.7991452991452992, "acc_norm_stderr": 0.026246772946890477 }, "harness|hendrycksTest-medical_genetics|5": { "acc": 0.63, "acc_stderr": 0.04852365870939099, "acc_norm": 0.63, "acc_norm_stderr": 0.04852365870939099 }, "harness|hendrycksTest-miscellaneous|5": { "acc": 0.7535121328224776, "acc_stderr": 0.015411308769686934, "acc_norm": 0.7535121328224776, "acc_norm_stderr": 0.015411308769686934 }, "harness|hendrycksTest-moral_disputes|5": { "acc": 0.6445086705202312, "acc_stderr": 0.025770292082977254, "acc_norm": 0.6445086705202312, "acc_norm_stderr": 0.025770292082977254 }, "harness|hendrycksTest-moral_scenarios|5": { "acc": 0.42681564245810055, "acc_stderr": 0.016542401954631917, "acc_norm": 0.42681564245810055, "acc_norm_stderr": 0.016542401954631917 }, "harness|hendrycksTest-nutrition|5": { "acc": 0.5915032679738562, "acc_stderr": 0.028146405993096358, "acc_norm": 0.5915032679738562, "acc_norm_stderr": 0.028146405993096358 }, "harness|hendrycksTest-philosophy|5": { "acc": 0.6784565916398714, "acc_stderr": 0.026527724079528872, "acc_norm": 0.6784565916398714, "acc_norm_stderr": 0.026527724079528872 }, "harness|hendrycksTest-prehistory|5": { "acc": 0.654320987654321, "acc_stderr": 0.02646248777700187, "acc_norm": 0.654320987654321, "acc_norm_stderr": 0.02646248777700187 }, "harness|hendrycksTest-professional_accounting|5": { "acc": 0.44680851063829785, "acc_stderr": 0.029658235097666907, "acc_norm": 0.44680851063829785, "acc_norm_stderr": 0.029658235097666907 }, "harness|hendrycksTest-professional_law|5": { "acc": 0.4445893089960887, "acc_stderr": 0.012691575792657114, "acc_norm": 0.4445893089960887, "acc_norm_stderr": 0.012691575792657114 }, "harness|hendrycksTest-professional_medicine|5": { "acc": 0.5441176470588235, "acc_stderr": 0.030254372573976715, "acc_norm": 0.5441176470588235, "acc_norm_stderr": 0.030254372573976715 }, "harness|hendrycksTest-professional_psychology|5": { "acc": 0.5898692810457516, "acc_stderr": 0.019898412717635906, "acc_norm": 0.5898692810457516, "acc_norm_stderr": 0.019898412717635906 }, "harness|hendrycksTest-public_relations|5": { "acc": 0.5909090909090909, "acc_stderr": 0.047093069786618966, "acc_norm": 0.5909090909090909, "acc_norm_stderr": 0.047093069786618966 }, "harness|hendrycksTest-security_studies|5": { "acc": 0.6408163265306123, "acc_stderr": 0.030713560455108493, "acc_norm": 0.6408163265306123, "acc_norm_stderr": 0.030713560455108493 }, "harness|hendrycksTest-sociology|5": { "acc": 0.7661691542288557, "acc_stderr": 0.02992941540834839, "acc_norm": 0.7661691542288557, "acc_norm_stderr": 0.02992941540834839 }, "harness|hendrycksTest-us_foreign_policy|5": { "acc": 0.81, "acc_stderr": 0.039427724440366255, "acc_norm": 0.81, "acc_norm_stderr": 0.039427724440366255 }, "harness|hendrycksTest-virology|5": { "acc": 0.43373493975903615, "acc_stderr": 0.038581589406855174, "acc_norm": 0.43373493975903615, "acc_norm_stderr": 0.038581589406855174 }, "harness|hendrycksTest-world_religions|5": { "acc": 0.8070175438596491, "acc_stderr": 0.030267457554898458, "acc_norm": 0.8070175438596491, "acc_norm_stderr": 0.030267457554898458 }, "harness|truthfulqa:mc|0": { "mc1": 0.31334149326805383, "mc1_stderr": 0.016238065069059605, "mc2": 0.4605506661658282, "mc2_stderr": 0.014802420782627305 }, "harness|winogrande|5": { "acc": 0.7663772691397001, "acc_stderr": 0.011892194477183525 }, "harness|gsm8k|5": { "acc": 0.01288855193328279, "acc_stderr": 0.003106901266499642 } }