🚀 𝐓𝐇𝐄 𝐅𝐈𝐑𝐒𝐓 𝐒𝐄𝐍𝐓𝐈𝐄𝐍𝐓 𝐌𝐎𝐃𝐄𝐋, 𝐃𝐎𝐁𝐁𝐘-𝐦𝐢𝐧𝐢, 𝐈𝐒 𝐋𝐀𝐔𝐍𝐂𝐇𝐈𝐍𝐆 𝐓𝐎𝐃𝐀𝐘 The signature fingerprinted Dobby 70B model is coming in early February. 𝐃𝐨𝐛𝐛𝐲-𝐦𝐢𝐧𝐢 𝐜𝐨𝐦𝐞𝐬 𝐢𝐧 𝐭𝐰𝐨 𝐟𝐥𝐚𝐯𝐨𝐫𝐬 😇𝐃𝐨𝐛𝐛𝐲-𝐦𝐢𝐧𝐢 𝐋𝐞𝐚𝐬𝐡𝐞𝐝—Friendly and conversational, your regular chill guy 😈𝐃𝐨𝐛𝐛𝐲-𝐦𝐢𝐧𝐢 𝐔𝐧𝐡𝐢𝐧𝐠𝐞𝐝—Brutally honest and unfiltered, speaks uncomfortable truths others won't Try them now by running Dobby-mini... locally: 👉 …with a helpful how-to guide from our friend Chris Aubin: 👉show more

2/ 🚨𝐀𝐈 𝐄𝐍𝐆𝐈𝐍𝐄𝐄𝐑𝐈𝐍𝐆 𝐁𝐑𝐄𝐀𝐊𝐓𝐇𝐑𝐎𝐔𝐆𝐇: 𝐔𝐍𝐋𝐎𝐂𝐊𝐈𝐍𝐆 𝐏𝐑𝐄𝐂𝐈𝐒𝐄 𝐌𝐎𝐃𝐄𝐋 𝐀𝐋𝐈𝐆𝐍𝐌𝐄𝐍𝐓🚨 𝐒𝐞𝐧𝐭𝐢𝐞𝐧𝐭 𝐡𝐚𝐬 𝐝𝐞𝐯𝐞𝐥𝐨𝐩𝐞𝐝 𝐭𝐞𝐜𝐡𝐧𝐢𝐪𝐮𝐞𝐬 𝐭𝐨 𝐜𝐚𝐥𝐢𝐛𝐫𝐚𝐭𝐞 𝐀𝐈 𝐦𝐨𝐝𝐞𝐥𝐬 𝐰𝐢𝐭𝐡 𝐞𝐱𝐜𝐞𝐩𝐭𝐢𝐨𝐧𝐚𝐥 𝐠𝐫𝐚𝐧𝐮𝐥𝐚𝐫𝐢𝐭𝐲 𝐟𝐨𝐫 𝐭𝐡𝐞 𝐟𝐢𝐫𝐬𝐭 𝐭𝐢𝐦𝐞 𝐞𝐯𝐞𝐫. Imagine an AI that can be precisely tuned to ANY worldview or value system without compromising core safety protocols. Sentient’s "Loyalty Training" methodology allows precise model alignment: Want a model that aligns with conservative political views? We can do it. Want a model that aligns with protecting the environment? We can do it. On the Sorry Bench, one can observe how both Dobby-mini models have been meticulously aligned to specific biases while still maintaining the essential safety benchmarks. 𝐓𝐇𝐈𝐒. 𝐂𝐇𝐀𝐍𝐆𝐄𝐒. 𝐄𝐕𝐄𝐑𝐘𝐓𝐇𝐈𝐍𝐆.

3/ 𝐇𝐎𝐖 𝐃𝐎𝐄𝐒 𝐃𝐎𝐁𝐁𝐘-𝐦𝐢𝐧𝐢 𝐂𝐎𝐌𝐏𝐀𝐑𝐄 𝐓𝐎 𝐎𝐓𝐇𝐄𝐑 𝐌𝐎𝐃𝐄𝐋𝐒 𝐖𝐞 𝐚𝐫𝐞 𝐭𝐡𝐞 𝐟𝐢𝐫𝐬𝐭 𝐭𝐞𝐚𝐦 𝐭𝐨 𝐢𝐥𝐥𝐮𝐬𝐭𝐫𝐚𝐭𝐞 𝐚𝐥𝐢𝐠𝐧𝐦𝐞𝐧𝐭 𝐀𝐍𝐃 𝐫𝐞𝐭𝐞𝐧𝐭𝐢𝐨𝐧 𝐨𝐟 𝐦𝐨𝐝𝐞𝐥 𝐮𝐭𝐢𝐥𝐢𝐭𝐲. Fine-tuning models for significant changes can devolve a model into complete nonsense. Compared with Llama 8B, Dobby-mini retains the performance of the underlying Llama model on key benchmarks such as IFEval and Math-hard. This is remarkable since significantly fine-tuning models results in a “seesaw” effect where improvement in one eval results in a drop in another. For instance, Hermes 3 8B was also trained on Llama 8B but dropped significantly in performance on these crucial benchmarks, severely limiting the model’s utility in instruction following and reasoning. Dobby is entertaining (through tone change) and retains its utility at par with the original Llama 8B.

4/𝐖𝐇𝐘 𝐈𝐒 𝐓𝐇𝐈𝐒 𝐈𝐌𝐏𝐎𝐑𝐓𝐀𝐍𝐓 𝐒𝐞𝐧𝐭𝐢𝐞𝐧𝐭’𝐬 𝐀𝐈 𝐑𝐞𝐬𝐞𝐚𝐫𝐜𝐡 𝐭𝐞𝐚𝐦 𝐡𝐚𝐬 𝐪𝐮𝐢𝐜𝐤𝐥𝐲 𝐛𝐞𝐜𝐨𝐦𝐞 𝐭𝐡𝐞 𝐛𝐞𝐬𝐭 𝐀𝐈 𝐭𝐞𝐚𝐦 𝐢𝐧 𝐭𝐡𝐞 𝐜𝐫𝐲𝐩𝐭𝐨 𝐢𝐧𝐝𝐮𝐬𝐭𝐫𝐲, with deep expertise in post-training. The team has solved two problems: 𝟏. 𝐇𝐨𝐰 𝐝𝐨 𝐲𝐨𝐮 𝐭𝐮𝐧𝐞 𝐚 𝐦𝐨𝐝𝐞𝐥’𝐬 𝐚𝐥𝐢𝐠𝐧𝐦𝐞𝐧𝐭 𝐨𝐧 𝐬𝐩𝐞𝐜𝐢𝐟𝐢𝐜 𝐝𝐢𝐦𝐞𝐧𝐬𝐢𝐨𝐧𝐬? 𝟐. 𝐇𝐨𝐰 𝐝𝐨 𝐲𝐨𝐮 𝐞𝐧𝐬𝐮𝐫𝐞 𝐭𝐡𝐢𝐬 𝐚𝐥𝐢𝐠𝐧𝐦𝐞𝐧𝐭 𝐝𝐨𝐞𝐬 𝐧𝐨𝐭 𝐝𝐞𝐠𝐫𝐚𝐝𝐞 𝐩𝐞𝐫𝐟𝐨𝐫𝐦𝐚𝐧𝐜𝐞? Dobby-mini illustrates the first step solution of these two research questions in the form of an entertaining model that is aligned to specific values (pro-crypto and pro-personal freedom), while still maintaining the same performance as Llama 3.1 8B.

5/ 𝐄𝐗𝐏𝐄𝐑𝐈𝐄𝐍𝐂𝐄 𝐓𝐇𝐄 𝐃𝐎𝐁𝐁𝐘-𝐦𝐢𝐧𝐢 𝐌𝐎𝐃𝐄𝐋𝐒 𝐒𝐞𝐞 𝐡𝐨𝐰 𝐨𝐭𝐡𝐞𝐫𝐬 𝐮𝐬𝐞 𝐃𝐨𝐛𝐛𝐲-𝐦𝐢𝐧𝐢 by joining our Discord channel: 👉 𝐓𝐚𝐥𝐤 𝐭𝐨 𝐃𝐨𝐛𝐛𝐲-𝐦𝐢𝐧𝐢 through an interactive demo run by Gradio with limited capacity (Please note this is a soft demo not built for retail scale yet, a scaled version will be coming soon😁) 👉 𝐁𝐮𝐢𝐥𝐝 𝐨𝐧 𝐭𝐨𝐩 𝐨𝐟 𝐃𝐨𝐛𝐛𝐲-𝐦𝐢𝐧𝐢 through the API provided by our friends @FireworksAI_HQ: 👉

6/ 𝐅𝐎𝐋𝐋𝐎𝐖 𝐎𝐔𝐑 𝐂𝐎𝐑𝐄 𝐂𝐎𝐍𝐓𝐑𝐈𝐁𝐔𝐓𝐎𝐑𝐒 Our friends will be posting some sneak peeks and extra info on Dobby: @sandeepnailwal @hstyagi @oleg_golev @abhishek095 @vivekkolli @antoniok1406 @crestonbrooks @ee_mc3 @cvongerlach

Step into comfort and style with our newly arrived men's jumpsuits and two-piece sets. Upgrade your wardrobe and make every occasion a runway!

cheeky dobby

My first confersation with DOBBY-mini model B I thought they were just jokeing when they say it's honest I think they need to add +18 in the next time when say it's honest 😬

I will test it 🫡

🤝