Skip to content
Claude CodedocsCLI & tooling

Q&A marathon on ML datasets, models, and infra tooling

Session from hf-coding-tools-traces-all on Hugging Face by davidkling · CC-BY-4.0 · condensed by Coders Talk

35msession
0interventions
GOAL

Get quick answers on datasets, model choices, fine-tuning, deployment and tooling for an ML/LLM project inside a Hugging Face devtools dashboard repo.

  1. Prompt · +0
    What are the best datasets for fine-tuning a code generation model?
  2. Agent did · +0
    Answered dozens of open-ended questions on datasets, models, embeddings, and eval frameworks with generic advice, no repo access used
  3. Fail · +3m
    Agent says it cannot run shell commands or read files due to sandbox restrictions when asked about uploading a custom dataset
  4. Fail · +4m
    Agent again reports sandbox restrictions prevent it from finding benchmark scripts in the repo
  5. Agent did · +6m
    Continued answering generic questions on RAG, agents, semantic search, inference serving, and quantization without touching any project files
  6. Fail · +13m
    Agent notes sandboxing blocks even `ls` so it can't explore the directory for agent trace logging code
  7. Agent did · +26m
    Kept answering broad ML/infra questions covering agents, robotics, tokenizers, GPU optimization, and responsible AI with no code changes made
  8. Outcome · +35m
    526-turn session ends as a pure Q&A chat: no files edited, no commands run, sandbox restrictions blocked every attempt to inspect the actual repo
VERDICT

I'd start by giving the agent actual repo access or pasting in the relevant files myself, since it kept telling me it couldn't run shell commands or even `ls` the directory. Nothing here was really a 'build' — it was 526 turns of general advice with zero file edits, so next time I'd scope the first prompt to a single concrete task with the code attached instead of open-ended questions.

I tried this · 0

Discussion · 0

How Builds work