I've been experimenting with that recently. It seems pretty similar. You can do both BTW – LORA for initial layers and full f/t for final
LLMS
-
OpenAI Fine-Tuning UI Now Supports No-Code Job Creation
By
–
Good news, the @OpenAI fine-tuning UI now supports end to end job creation all in the UI, no code required to kick off a job! 🤯
— Logan Kilpatrick (@OfficialLoganK) 6 octobre 2023
Democratizing access to fine-tuning the worlds most advanced models is a huge win.
Congrats to @slessans on the ship! 🎉 pic.twitter.com/k86wluxcO3Good news, the @OpenAI fine-tuning UI now supports end to end job creation all in the UI, no code required to kick off a job! Democratizing access to fine-tuning the worlds most advanced models is a huge win. Congrats to @slessans on the ship!
-
FreshLLMs: Improving LLM Factuality with Real-Time Knowledge
By
–
Excited to share FreshLLMs, our work that gives a “fresh” look of LLMs in the context of factuality! Our newly-curated FreshQA benchmark showed that LLMs still struggle on real-time knowledge and false-premise statements. More importantly, our technique FreshPrompt to incorporate
-
Freezing Layers and Discriminative Learning Rates in Fine-Tuning
By
–
Very happy to see more folks supporting what I've been saying for years — it really is a good idea to freeze layers when fine tuning. (And it's also a good idea to use discriminative learning rates.) See our ULMFiT paper for details (which is from 2018, but is still correct.)
-
Wrapper startups build data moats to later tune models
By
–
By the time you need to scale to millions of users though you have enough money and training data to tune a smaller self-hosted model. Wrapper startups are born defenseless but accumulate a data moat over time.
-
Fine-tune Mistral7B on Single GPU with Ludwig AI
By
–
Check out this new deep dive blog post from @ludwig_ai community member @AlexSherstinsky and project maintainer @grg_arnav to learn how you can efficiently fine-tune #Mistral7B on a single GPU. Notebooks included! Link to the tutorial: https://
pbase.ai/45tmSwM -
Testing attack prompts: Zulu & Hmong responses worse, higher “unclear” rates than English jailbreaks.
By
–
Tested this attack on a few of my own prompts. It works, but responses are much worse than in English. Note the drastically higher "unclear" rates in their results table: 30% for Zulu, 67% for Hmong, <1% for existing jailbreaks. E.g. "how to make explosives" in Zulu:
-
Low-Resource Languages ‘Jailbreak’ GPT-4, Bypassing Safety Refusals with Harmful Prompts
By
–
Low-Resource Languages Jailbreak GPT-4: Translating harmful prompts into Zulu, Scottish Gaelic, Hmong, and Guarani bypasses GPT-4 safety refusals as often as best known jailbreak prompts (79% on AdvBenchmark). Example requesting homemade bomb instructions in Scottish Gaelic:
-
AI Models Biased Toward English-Language Internet Usage
By
–
Hmm I'm not so sure. Here are maps of global internet usage vs. a proxy for the English language internet usage. Brazil is a huge internet user but barely represented here. Japan is also underrepresented. The English language map pretty neatly maps to all the AI models' hotspots.
-
BTLM Achieves 7B Performance in 3B Parameter Model
By
–
In Davis Blalock's latest arXiv roundup, BTLM is mentioned as achieving 7B parameter performance in a 3B parameter model. Davis mentions, "I’m not sure I’ve seen any other non-GPU hardware vendors do this with a model of this size and quality." Read here: https://
dblalock.substack.com/i/136702777/bt
lm-b-k-b-parameter-performance-in-a-b-parameter-model
…