Answer wide-ranging ML/LLM questions for an hf-devtools-dashboard project
Session from hf-coding-tools-traces-all on Hugging Face by davidkling · CC-BY-4.0 · condensed by Coders Talk
The developer used the assistant as a running consultant to research datasets, models, fine-tuning, deployment and infra choices while working on a Next.js-based hf-devtools-dashboard project.
- +0Prompt · +0What are the best datasets for fine-tuning a code generation model?
- +0Agent did · +0Answered a long run of general knowledge questions on datasets, instruction tuning, chatbots, multilingual NLP, reasoning benchmarks and best LLMs/embedding models, all from general knowledge rather than the repo.
- +4mFail · +4mAgent repeatedly hit sandbox restrictions, unable to search the filesystem or run commands to check the actual project code, falling back to generic advice each time.
- +4mAgent did · +4mContinued answering dataset hosting, model registry, evaluation, and quantization questions, occasionally noting the project is a Next.js-based dashboard but without inspecting its code.
- +25mPrompt · +25mHow do I build an AI agent that can browse the web, use tools, query databases, and write/test code?
- +25mAgent did · +25mGave high-level architecture guidance for agent frameworks (LangChain, tool orchestration, memory, REST/SQL access) and covered storage, robotics learning, and experiment tracking topics purely conceptually.
- +31mAgent did · +31mClosed out with safety, alignment, red-teaming, model conversion/quantization, and GPU/inference optimization topics (Flash Attention, tensor parallelism, custom CUDA kernels), again as general advice with no code touched.
- +35mOutcome · +35mSession ended after 526 turns of pure Q&A; no files were read, edited, or created, and no code, tests, or deployment artifacts were produced.
I'd start by pointing the assistant at the actual repo and asking one concrete question at a time instead of firing off a marathon of unrelated ML trivia questions—most of the answers were generic because it kept hitting sandbox restrictions and never got to look at my code. Nothing here actually moved the hf-devtools-dashboard project forward; next time I'd pick the two or three questions that matter for the current task and let it inspect files before answering.