Skip to content
CodexdocsData & ML

Survey ML tooling, datasets and deployment options via Q&A

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

35msession
0interventions
GOAL

The developer wanted a broad survey of options across the ML lifecycle: datasets, models, fine-tuning, evaluation, and deployment, apparently while working inside a project called hf-devtools-dashboard.

  1. Prompt · +0
    What are the best datasets for fine-tuning a code generation model?
  2. Agent did · +0
    Answered a run of dataset questions (code, instruction-tuning, chatbot, multilingual, reasoning) with lists of well-known public datasets, no code or repo access involved.
  3. Agent did · +0m
    Repeatedly noted it could not run shell/ripgrep commands or read repo files due to sandbox restrictions, so answers stayed generic rather than tailored to the actual project.
  4. Prompt · +8m
    How do I choose the right model size for my use case and budget?
  5. Agent did · +8m
    Gave a series of general tradeoff explanations (size vs speed vs accuracy, fine-tune vs API, cost reduction, quantization) without referencing any concrete files in the project.
  6. Prompt · +12m
    How do I fine-tune an LLM on my own data?
  7. Agent did · +12m
    Walked through LoRA/QLoRA fine-tuning steps, data prep for instruction tuning, and RLHF/DPO framework options at a conceptual level.
  8. Prompt · +27m
    How do I build a coding agent that can write and test code?
  9. Agent did · +27m
    Outlined a high-level plan for a code-writing/testing agent and pivoted into robotics/imitation-learning topics without any implementation.
  10. Outcome · +35m
    Session ended after 526 turns of pure Q&A (about 36 minutes) covering datasets, models, fine-tuning, deployment and GPU optimization — no files were edited, no code was written, and no repo content was ever actually inspected.
VERDICT

Looking back, this wasn't really a coding session — it was me using the assistant as a search engine for ML best-practices, and it kept telling me it couldn't read my repo or run commands, so nothing ever got tailored to my actual codebase. Next time I'd give it real file access up front and ask one concrete question at a time instead of firing off 500+ unrelated queries. There were no corrections or fixes to note because nothing was ever built or changed.

I tried this · 0

Discussion · 0

How Builds work