Q&A on ML/LLM tooling, datasets, and infra topics
Session from hf-coding-tools-traces-all on Hugging Face by davidkling · CC-BY-4.0 · condensed by Coders Talk
The developer asked a broad, unstructured series of questions about datasets, models, fine-tuning, deployment, and infra topics for ML/LLM projects, seemingly probing general knowledge rather than working on a specific codebase task.
- +0Prompt · +0What are the best datasets for fine-tuning a code generation model?
- +0Agent did · +0Answered a long sequence of general knowledge questions about datasets, chatbots, multilingual NLP, and top LLMs for various tasks, with no reference to any actual repo or codebase.
- +3mAgent did · +3mStarted referencing a 'Next.js-based project' (hf-devtools-dashboard) and AGENTS.md in answers, despite the conversation still being purely Q&A with no file edits or commands run.
- +4mFail · +4mRepeatedly stated it could not run search or file commands due to 'sandbox restrictions', unable to actually inspect the codebase it kept referencing.
- +12mAgent did · +12mAnswered many fine-tuning method questions (LoRA, QLoRA, DPO, RLHF, distillation) purely as generic advice, no code written or tested.
- +25mAgent did · +25mContinued answering high-level questions about building AI agents, robotics learning, and orchestration frameworks, still without any concrete implementation.
- +31mAgent did · +31mCovered safety/alignment topics (guardrails, RLHF alignment, red-teaming) as general advice.
- +35mOutcome · +35mSession ended after 526 turns of pure Q&A; no code was written, no files edited, and no tests run — the agent never got to act on the referenced project.
I never actually pointed this at a real task — I just fired off question after question and never gave the agent a concrete file or repo to work with, so it kept guessing at a 'Next.js project' it couldn't inspect. Next time I'd start by pointing it at the actual repo and asking for one specific change instead of treating it like a search engine for two hours. The one thing that would have helped was letting it run file searches instead of hitting sandbox restrictions every time it tried to ground an answer in real code.