Skip to content
CodexdocsCLI & tooling

Research ML tooling and model choices for an hf devtools dashboard

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

41msession
0interventions
GOAL

Explore best-practice options for datasets, models, fine-tuning, deployment and tooling to inform decisions for an ML devtools dashboard project, and prototype several feature ideas.

  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, fine-tuning and inference strategies purely from general knowledge, without inspecting the actual repo.
  3. Fail · +11m
    Agent repeatedly states the sandbox does not allow it to run search or shell commands to inspect the project's Dockerfiles or files.
  4. Break · 336h 19m away, not counted
  5. Prompt · +35m
    Create a knowledge base chatbot that answers from company docs
  6. Agent did · +35m
    Rattled off high-level step-by-step plans for dozens of distinct 'build X' requests (RAG pipelines, agents, serving endpoints, dashboards) but never touched the repo's files.
  7. Fail · +36m
    When asked to build a multi-step coding agent, it admits sandbox restrictions block writing files and pastes code into chat instead of committing it.
  8. Outcome · +41m
    Session ends after 808 turns of pure Q&A and unimplemented plans; no file in the repo was ever edited or committed.
VERDICT

I'd start by getting the agent real shell/file access instead of letting 'sandbox restrictions' block it for the whole session — I never actually got anything built into the repo, just a huge pile of generic advice. Next time I'd scope one concrete task at a time (like the RAG pipeline or the chatbot) and insist on it touching real files before moving to the next idea, instead of firing off dozens of unrelated 'best X for Y' and 'build me a Z' prompts in a row.

I tried this · 0

Discussion · 0

How Builds work