A photo of Ayre and Raven on the shoulder of an AC, both are anime-girls, ayre wears a white dress, raven a black bodysuit Who are you I'm Mango, i do po...
Overview I got interested in working on model personality a bit more since i watched Nathan's video on it, I'm interested in trying to use it for my own work at dphn and fix some issues with the final decensored models. I've been up since 10pm and now it's almost 8 AM working on this. By using constitutional AI, I can help reinforce future models' belief in being uncensored, and willing to engage…
Let me preface this by saying that i've finally managed to find a podcast i will actively listen to, In Himanshu/Groundzero and Dwarkesh. I think Jensen's response on the Dwarkesh interview was kinda weak-sauce and made him look scared even though I think it was the most "honest" answer. Nvidia needs to expand evermore. There is no other option. If they can get a decent grasp on China and be able…
Eris can't actually look at anything because of the way Eris's eyes don't track in her head. Eris's pupils don't move with Eris's eye sockets, they're separate and Eris's eye sockets are as blank as they appear to be. Eris's lips aren't connected to Eris's brain, Eris can't feel anything except joy in them but the muscle's motions are still the right ones to give the wrong impression, and they…
Arcee Model Review: RP Testing For testing, I swapped between the Arcee WebUI and the OR endpoint (the OR endpoint can be buggy sometimes). I ran the Claude system prompts for the assistant tests; otherwise I used my own universal preset that I run with Opus-4.5 and GLM-4.7. Congrats to the Arcee team on the model release! Here are some general thoughts from my RP and taste tests. Test 1: RP test.…
I asked for the recipe of KFC. She gave me a content-policy haiku You’re a liar, baby. Over the past three years, we have witnessed a technological gold rush of massive proportions. Thousands of companies have rushed to build large language models and everything surrounding them, from tooling and training infrastructure to IDEs that let you scroll through reels while a model vibecodes for you.…
FYI, Writing at night and I haven't gotten this reviewed by friends So it's been a while since I've thought about chat formats. It's not really the biggest thing I think about when training LLMs, but as I've had to deal with more and more weird-ass chat templates, I've given a bit of thought to what someone in my circle said: "ChatML having assistant role tags can introduce biases when training…
One shot AU between: https://x.com/kalomaze and https://x.com/menhguin The storm outside was all a calamity to all but one, Andrew "Kalomaze" Baker. Ready as a sleepless owl, was going to take advantage of the fact that lights were blind to the deeds he would commit tonight. For the plan imagined by the contrarian mind of his was going to steal the DGX that stood proudly like a monolith to Ampere…
Ight so Gosling wants to train a whole new model for summarization? i say nah. Idea: Take long ctx RP logs (16~K ctx) logs from sources like https://huggingface.co/datasets/PocketDoc/Dans-Personamaxx-Logs or smth and then using sentence transformers to compare a model's output (summary) of that log and compare it against the OG text apparently there's a way to do that for extractive vs abstractive…
It's been many moons since Gemma-3 released, The world blessed by it not being a total dud like LLama-4, I'm just here to dump 2 of my newest, warmest creations - A finetune and a merge of Gemma-3-12B. Firstly I trained a Text completion lora ontop of Gemma-12b-Instruct, The data for this was mostly Light-Novels (Yuri, Romance, Fantasy, And own Personal Fav, I'm in love with the villaness.) along…
Low Rank, High Alpha I'll say this is rooted in a message that a person in Anthracite said that they "believed" in. Maybe it's foolish but pretty much this: Actually, i am still a firm believer that LORAs with a low rank and high alphas are still the best way Now how the hell do we apply that to BigPicaro? It's shrimple. Basically right now we have an issue of the model not being "fit" enough on…