[{"data":1,"prerenderedAt":183},["ShallowReactive",2],{"project-43846690":3},{"id":4,"name":5,"line":6,"line_source":7,"title":8,"url":9,"domain":10,"maker":11,"posted_at":12,"year":13,"points":14,"comments":15,"rank_in_year":16,"shown_in_year":17,"top":18,"tier":19,"category":20,"subcategory":21,"format":22,"keywords":23,"reception":28,"sentiment":29,"stance_mix":30,"scored_comments":34,"aspects":35,"strip":40,"post_text":41,"what":42,"said":58,"maker_in_thread":100,"shape":107,"grid":110,"attempts":145,"same_week":146,"maker_shows":31,"maker_hidden":175,"reader_version":176,"card":177},"43846690","ART","a new open-source RL framework for training agents","title","Show HN: ART – a new open-source RL framework for training agents","https:\u002F\u002Fgithub.com\u002FOpenPipe\u002FART","github.com","kcorbitt",1746027300,2025,116,12,647,3837,17,"notable","ai","reinforcement learning framework","library",[24,25,26,27],"reinforcement learning","llm agents","agent training","openai api","loved",0.479,{"positive":31,"neutral":32,"mixed":33,"critical":33},5,3,0,8,{"design":36,"idea":38,"docs":39},{"positive":37,"critical":33},1,{"positive":37,"critical":33},{"positive":32,"critical":33},[31,33,32,33,31,31,32,31,33,33,31,32],"Hey HN, I wanted to share a new project we've been working on for the last couple of months called ART (https:\u002F\u002Fgithub.com\u002FOpenPipe\u002FART).\n\nART is a new open-source framework for training agents using reinforcement learning (RL). RL allows you to train an agent to perform better at any task whose outcome can be measured and quantified.\n\nThere are many excellent projects focused on training LLMs with RL, such as GRPOTrainer (https:\u002F\u002Fhuggingface.co\u002Fdocs\u002Ftrl\u002Fmain\u002Fen\u002Fgrpo_trainer) and verl (https:\u002F\u002Fgithub.com\u002Fvolcengine\u002Fverl). We've used these frameworks extensively for customer-facing projects at ",{"problem":43,"audience":49,"pricing":49,"motivation":50,"time_to_build":49,"stack":54,"openness":55,"stage":56,"team_size":57},{"value":44,"source":45,"evidence":46},"Training agents with RL is hard for multi-turn tool-calling workflows and GPU efficiency is low in existing trainers","maker",[47],{"id":48},"post",null,{"value":51,"source":45,"evidence":52},"Existing RL frameworks lacked multi-turn support, GPU efficiency and chat completion APIs for agent codebases",[53],{"id":48},[],"open_source","launched","company",{"praise":59,"objections":74,"requests":75,"questions":83,"alternatives":91},[60,67],{"value":61,"source":62,"evidence":63},"Clear explanation of SFT versus RL","commenter",[64],{"id":65,"by":11,"text":66},"43849667","Ok good questions here. By fine-tuning in this context I assume you mean \"supervised fine-tuning\", or SFT. SFT trains a model to produce a specific string of output tokens, given…",{"value":68,"source":62,"evidence":69},"Useful documentation and comparison table",[70],{"id":71,"by":72,"text":73},"43851229","jeffchuber","the table with comparable models is a really great way to show off things here",[],[76],{"value":77,"source":62,"evidence":78},"Documentation for HTTP API endpoints such as \u002F_train_model",[79],{"id":80,"by":81,"text":82},"43849296","tcdent","I really like this concept. Do you have documentation for the API response from the `\u002F_train_model` endpoint?",[84],{"value":85,"source":62,"evidence":86},"How does RL differ from fine tuning and when to use each",[87],{"id":88,"by":89,"text":90},"43849331","someguy101010","Thanks for sharing this! A couple of questions come to mind: - How does training with RL differ from fine tuning? - When would it make sense to fine tune instead of using RL?",[92,96],{"value":93,"source":45,"evidence":94},"GRPOTrainer",[95],{"id":48},{"value":97,"source":45,"evidence":98},"verl",[99],{"id":48},{"comments":101,"answered":37,"key":102},2,[103,106],{"id":104,"text":105},"43848552","Figured now was a good time to post this since we recently got surprisingly good results on training an email research agent. Link is above, but will put it here as well since I…",{"id":65,"text":66},{"commenters":34,"max_depth":32,"maker_replies":101,"first_comment_mins":108,"lifespan_hours":109},132,19,[111,115,117,119,120,123,125,128,130,134,137,142],[112,33,31,113,114,33],"43848831","project\u002Fpositive\u002Fdesign","bradhilton",[104,37,33,116,11,37],"project\u002Fpositive\u002Fperformance",[88,101,32,118,89,33],"project\u002Fneutral\u002Fidea",[65,101,33,118,11,37],[121,101,31,122,89,33],"43850506","maker\u002Fpositive\u002Fdocs",[80,32,31,124,81,33],"project\u002Fpositive\u002Fdocs",[126,32,32,127,114,33],"43849385","project\u002Fneutral\u002Fdocs",[71,129,31,124,72,33],4,[131,31,33,132,133,33],"43854214","other\u002Fneutral\u002Fnone","pama",[135,31,33,132,136,33],"43854401","schainks",[138,139,31,140,141,33],"43855666",6,"project\u002Fpositive\u002Fidea","gitroom",[143,139,32,144,114,33],"43856094","maker\u002Fneutral\u002Fnone",[],[147,152,157,162,167,171],{"id":148,"name":149,"points":150,"category":151},"43820228","PyXL",983,"hardware",{"id":153,"name":154,"points":155,"category":156},"43803148","RetrOS-32",575,"infrastructure",{"id":158,"name":159,"points":160,"category":161},"43816419","screenrecorder.me",462,"dev_tools",{"id":163,"name":164,"points":165,"category":166},"43835584","Beatsync",439,"creative",{"id":168,"name":169,"points":170,"category":156},"43778081","TacOS",320,{"id":172,"name":173,"points":174,"category":166},"43791992","Clever Coloring Book",316,false,"v2",{"id":178,"name":5,"line":6,"ai":175,"cat":20,"year":13,"points":14,"comments":15,"rec":28,"s":179,"top":18,"kw":180,"maker":11,"domain":10,"tier":19,"attempts":33,"strip":181,"url":9,"mix":182,"praise":61,"objection":-1,"sub":21},43846690,0.48,[24,25,26,27],"503055350053",[31,32,33,33],1794036673932]