Top all-time · #3524
Quixi AI
Personal blog · erichartford.comai
Quixi AI is a personal blog at erichartford.com about ai. Since 2023, 22 different Hacker News readers have linked it 24 times, most often “Uncensored (Language) Models”. It is still writing: the last post was on 2026-08-29, and you can follow it by feed.
Too few sampled comments (15) to say how often HN recommends it.?
Why HN links it
HN readers link Eric Hartford's blog for its guide to uncensored LLMs and fine-tuning, which commenters call a good writeup and a great read. Other posts cover the Samantha and Dolphin models.
Summarised by Claude from a sample of 15 HN comments that link here. Each comment is marked recommending, critical or neutral (citing without judging); how this works.
Applied AI Researcher I make AI models like Dolphin and Samantha https://ko-fi.com/erichartford BTC 3ENBV6zdwyqieAXzZP2i3EjeZtVwEmAuo4 ETH 0xcac74542A7fF51E2fb03229A5d9D0717cB6d70C9 https://quixi.ai
Most-linked posts?
- Uncensored (Language) Models9 people · 9 mentions · first linked 2023-05-15
- /meet-samantha2 people · 3 mentions · first linked 2023-07-09
- From Zero to Fineturning with Axolotl on ROCm2 people · 2 mentions · first linked 2024-04-16
- Dolphin2 people · 2 mentions · first linked 2023-07-23
- Based-30B2 people · 2 mentions · first linked 2023-06-03
What Hacker News says?
"Dolphin, an emulator for Nintendo video games", but in this context it refers to "Dolphin, an open-source and uncensored, and commercially licensed dataset and series of instruct-tuned language models based on Microsoft's Orca paper." https://erichartford.com/dolphin
Comment linking /dolphin
on “ChatGPT for Teams”Submitted NIST's DeepSeek "evaluation" is a hit piece as a story · 278 points
HN discussionI hope your business works out for you and I am willing to believe that AMD has improved somewhat, but I do not believe AMD has improved enough to be worth people’s time when Nvidia is an option. I have heard too many nightmares and it is going to take many people, including people who reported those nightmares, reporting improvements for me to think otherwise. It is not just George Hotz who reported issues. Eric Ha…
Comment linking /from-zero-to-fineturning-with-axolotl-on-rocm
on “Why DeepSeek is cheap at scale but expensive to run locally”Submitted Uncensored (Language) Models as a story · 35 points
HN discussionOllama is not as powerful as llama.cpp or raw pytorch, but it is almost zero effort to get started. brew install ollama; ollama serve; ollama pull llama3: 8b-v2.9-q5_K_M; ollama run llama3: 8b-v2.9-q5_K_M https://ollama.com/library/dolphin-llama3:8b-v2.9-q5_K_M (It may need to be Q4 or Q3 instead of Q5 depending on how the RAM shakes out. But the Q5_K_M quantization (k-quantization is the term) is generally the best…
Comment linking /uncensored-models
on “Cost of self hosting Llama-3 8B-Instruct”Submitted Based-30B as a story · 3 points
HN discussion> properties like creativity, liveliness, and consistency I don't believe we train The LLMs for those things specifically. It will be interesting to see if some datasets for this appear. I think we can still make huge improvements by just caring about that more. We've got Samantha though, so I hope we will see those attributes too. https://erichartford.com/meet-samantha
Comment linking /meet-samantha
on “Does Reasoning Emerge? Probabilities of Causation in Large Language Models”Submitted Dolphin as a story · 2 points
HN discussionIn general, the core language model is simply trained on a very large amount of unannotated text (which is the most time-consuming and expensive part), but a language model is not directly very useful in the role of e.g. a chat agent, it quite literally tries to continue text and that sometimes is what you want and sometimes isn't. The second step is fine-tuning the model on a much smaller set of annotated data whic…
Comment linking /uncensored-models
on “Latest GPT-4 training data updated to December 2023”Submitted Uncensored Models (2023) as a story · 1 points
HN discussionedit: misread the title The uncensored one [1] - finally gave me instructions for making crack and a bomb. It felt cool that it would answer everything, like a 90s zine. [1] https://huggingface.co/TheBloke/dolphin-2.1-mistral-7B-GGUF Note on that page: This model is uncensored. I have filtered the dataset to remove alignment and bias. This makes the model more compliant. You are advised to implement your own alignme…
Comment linking /uncensored-models
on “Ask HN: What's the coolest non standard application of LLMs you've seen?”Submitted From Zero to Fineturning with Axolotl on ROCm as a story · 1 points
HN discussionCorrect. Eric Hartford's blog post delves into the alignment of open-source LLMs[1]. In essence, models like LLaMA and GPT-Neo-X adopt alignment behaviors from ChatGPT-sourced instruction datasets. To achieve more transparent model responses, one can refine the dataset by removing biases and refusals, then retrain. [1] https://erichartford.com/uncensored-models#heading-ok-so-if-...
Comment linking /uncensored-models
on “Comparing 60 LLMs with a set of 20 prompts”Submitted My Own AI Server Cluster as a story · 1 points
HN discussion1: https://erichartford.com/meet-samantha 2: ...Nope.
Comment linking /meet-samantha
on “Riley Reid on AI: 'I Don't Want Porn to Get Left Behind'”Submitted OpenOrca: open source dataset and instruct-tuned LLMs as a story · 1 points
HN discussionThis post seems to be upvoted for the "uncensored" keyword. But this should be attributed to https://huggingface.co/ehartford and others. See also https://news.ycombinator.com/item?id=36977146 Or better: https://erichartford.com/uncensored-models
Comment linking /uncensored-models
on “Run Llama 2 Uncensored Locally”Submitted Based-30B: a window into the mind of the foundational LLM model as a story · 1 points
HN discussionEric’s blog is a great read on how to create the uncensored models - link to the original blog here: https://erichartford.com/uncensored-models
Comment linking /uncensored-models
on “Run Llama 2 Uncensored Locally”Here's a good writeup that goes into more depth than most of the READMEs on fine-tuning: https://erichartford.com/uncensored-models
Comment linking /uncensored-models
on “XGen-7B, a new 7B foundational model trained on up to 8K length for 1.5T tokens”Try out Samantha, a llama 33b finetune, which may already be what you want: https://erichartford.com/meet-samantha Run this one on exLLaMA\tui if you happen to have a 3090/4090 to run it on: https://huggingface.co/TheBloke/Samantha-33B-SuperHOT-8K-GPT... And otherwise this on any decent PC: https://huggingface.co/TheBloke/Samantha-33B-SuperHOT-8K-GGM... That would be a good starting point for a finetune if its not q…
Comment linking /meet-samantha
on “How did you fine tune your chatbot?”I compiled this list - https://gist.github.com/TikkunCreation/5de1df7b24800cc05b482... In particular, you'll probably want to skip to nanoGPT ( https://github.com/karpathy/nanoGPT ) and then maybe if you are interested in a bit more of the theory, Zero to Hero ( https://karpathy.ai/zero-to-hero.html ), and his comments in one of the threads linked: https://news.ycombinator.com/item?id=34414716 Fine tuning may also b…
Comment linking /uncensored-models
on “Ask HN: Resource to learn how to train and use ML Models”I don't think so: https://erichartford.com/uncensored-models#heading-whats-an-... > Most of these models (for example, Alpaca, Vicuna, WizardLM, MPT-7B-Chat, Wizard-Vicuna, GPT4-X-Vicuna) have some sort of embedded alignment > The reason these models are aligned is that they are trained with data that was generated by ChatGPT, which itself is aligned by an alignment team at OpenAI.
Comment linking /uncensored-models
on “M2 Ultra can run 128 streams of Llama 2 7B in parallel”These are not the uncensored models. It's a fine tune of the censored models [1]. > Filter refusals and bias from the dataset -> finetune the model -> release. The alignment tax should still exist, maybe doubly so. [1] https://erichartford.com/uncensored-models#heading-lets-get-...
Comment linking /uncensored-models
on “M2 Ultra can run 128 streams of Llama 2 7B in parallel”
