Rapid-fire Q&A on ML/LLM tooling without a repo task
Session from hf-coding-tools-traces-all on Hugging Face by davidkling · CC-BY-4.0 · condensed by Coders Talk
41msession
0interventions
GOAL
The developer wanted quick answers across a huge range of ML/LLM topics — datasets, model choice, fine-tuning, serving, agents, optimization — and later tried to get the agent to actually build several ML features in the hf-devtools-dashboard repo.
- +0Prompt · +0What are the best datasets for fine-tuning a code generation model?
- +0Agent did · +0Answered dozens of unrelated ML/infra questions (datasets, models, embeddings, hosting, local inference) with generic knowledge, never touching the actual repo
- +11mFail · +11mAgent repeatedly says sandbox restrictions block it from running search or shell commands, so it can't inspect the repo and falls back to generic advice
- Break · 336h 19m away, not counted
- +35mPrompt · +35mSeries of build requests: a knowledge-base chatbot, a conversational search engine over PDFs, a citation-aware Q&A system, a RAG pipeline for internal docs
- +35mAgent did · +35mSketched high-level step-by-step plans for each requested build but produced no repo edits, only outlines and one pasted Python class in chat
- +36mFail · +36mHit sandbox restrictions again on file writes and searches, repeatedly asking the developer to paste code manually instead of inspecting or editing anything itself
- +38mPrompt · +38mMoved into integration asks (LLM in Flask API, CI/CD model deploys, model serving behind API gateway) — again without repo access
- +41mOutcome · +41mSession closed after 807 turns with zero files edited, zero commits, and no working feature — a long Q&A log rather than a completed build
VERDICT
Looking back, I never gave the agent one concrete task and stuck with it — I fired off hundreds of disconnected questions instead of picking a single build and pushing it through. The sandbox restrictions blocking file search and writes killed every real attempt at the actual builds I asked for near the end, and I should have fixed that access first instead of piling on more prompts.