Danau5tin

7 ms·

1.
by Danau5tin 2mo ago
Great question. Today the answer is that there is nothing stopping this, but you are right this is a 100% valid reward hacking possibility. No signs of it in the run I did, but for any future larger scale runs, it would be wise to dedicate
2.
by Danau5tin 2mo ago
Very good point. Sifting through noisy data and creating curated datasets could be of great value. Perhaps worth a project by itself!
3.
by Danau5tin 2mo ago
Very cool! Thank you
4.
by Danau5tin 2mo ago
Strangely, I did not. I was expecting it and looking out for it the whole time. At least I never say a warning!
5.
by Danau5tin 2mo ago
Yes, thank you. Typo!
6.
by Danau5tin 2mo ago
Yes exactly, and also there is auto-research by Karpathy, Fable-5 on parameter golf etc which point to the direction of highly-assisted / autonomous AI training as the future. Wild.
7.
by Danau5tin 2mo ago
I am also curious! The inner-RL-loop models are only trained once, then disgarded. But the outer-RL-loop model is trained on the same tasks over and over again. I imagine it would overfit after many more steps, but perhaps with a larger set
8.
by Danau5tin 2mo ago
Yes I do agree with this. I believe we are shifting from "make the model good" (prompt/context engineering, etc) to "define good for the model" (success criteria/rubrics). Over time I believe this will become i
9.
by Danau5tin 2mo ago
I chose the key technical decision and direction (such as the system architecture, the tasks to train on, the stack of Tinker, Prime-RL & Runpod - all of which I know well) etc. The problems it would do well on are training small agenti
10.
by Danau5tin 2mo ago
Yes I'd heard the labs (Anthropic mostly) speaking about LLMs training LLMs, so I wanted to make things a little more concrete and test it out myself! Essentially you are correct though, my framework allows the agent access to compute,
11.
by Danau5tin 2mo ago
Thank you!
12.
by Danau5tin 2mo ago
Mainly Fable, but It was me who wanted to emojis added hah. I also of course edited the README by hand (crazy I know), but the code is entirely fable
13.
by Danau5tin 2mo ago
I RL-trained an agent whose job is to write RL training jobs for smaller models, and open-sourced the whole thing. The setup is two nested RL loops: - Outer loop: the trainer agent (Qwen3.6-35B-A3B, LoRA) is handed a task spec ("teach
14.
Show HN: I RL-trained an agent that trains models with RL (for ~$1.3k)(github.com)
107 points by Danau5tin 2mo ago|49 comments
15.
Scaling Coding-Agent RL to 32x H100s. 160% Improvement on Stanford's TBench(github.com)
2 points by Danau5tin 11mo ago|1 comments
16.
by Danau5tin 11mo ago
My RL trained multi-agent-coding model Orca-Agent-v0.1-14B reached a 167% higher relative score than its base model on Stanford's TerminalBench. I've open sourced everything. *What I did:* - I trained a 14B orchestrator model to b
17.
Show HN: Multi-Agent-Coder Is #12 on Stanford's TBench. Beats Claude Code(github.com)
5 points by Danau5tin 1y ago|1 comments
18.
My weekend project accidentally beat Claude Code – #12 on Stanford's TBench(github.com)
2 points by Danau5tin 1y ago|2 comments
19.
by Danau5tin 1y ago
Hitting a million brick walls with multi-turn RL training isn't fun, so I thought I would try something new to climb Stanford's leaderboard for now! So this weekend I was just tinkering with multi-agent systems and... somehow ende
20.
by Danau5tin 1y ago
Exactly my first thought when I realised the cost! Currently LoRA is not supported by rLLM (The team told me they aim to support in next release), but it is certainly possible to port to verl directly or another RL framework for sure. I ju
21.
Show HN: Terminal-Bench-RL: Training long-horizon terminal agents with RL(github.com)
125 points by Danau5tin 1y ago|12 comments