Skip to content
Claude CodedocsCLI & tooling

Survey ML/AI tooling, models and datasets across 600+ Q&A turns

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

12h 59msession
0interventions
GOAL

The developer ran a long back-to-back series of general machine-learning questions (datasets, models, fine-tuning, deployment, safety, performance) to see how a coding assistant handles broad advisory queries alongside its awareness of the actual codebase it's attached to.

  1. Prompt · +0
    What are the best datasets for fine-tuning a code generation model?
  2. Agent did · +0
    Answered a run of dataset/model selection questions: pretraining corpora, instruction-tuning sets, chatbot data, multilingual NLP, reasoning benchmarks, top code-gen and multimodal models, self-hostable alternatives.
  3. Agent did · +53m
    Moved to application-building questions: adding summarization, RAG chatbots, LLM backend integration, Gradio demo UIs, semantic search and streaming chat responses.
  4. Agent did · +1h 39m
    Covered fine-tuning mechanics: full fine-tuning, LoRA/QLoRA on single GPUs, distributed training setups, and RLHF frameworks like TRL.
  5. Agent did · +2h 15m
    Answered evaluation questions: comparing against public leaderboards, measuring hallucination and toxicity, building custom eval tasks, reproducible benchmarking across hardware.
  6. Agent did · +2h 31m
    Walked through building tool-using and multi-step AI agents: tool-call loops, web-browsing agents, multi-agent orchestration, SQL agents, memory, and a coding agent that writes and tests code.
  7. Break · 10h 16m away, not counted
  8. Agent did · +5h 49m
    Fielded a cluster of robotics questions (manipulation policies, imitation learning frameworks like LeRobot, sim-to-real deployment) that the assistant flagged as outside the codebase's scope.
  9. Agent did · +9h 31m
    Addressed safety and alignment topics: guardrails, toxicity filtering, red-teaming, responsible deployment, RLHF alignment, bias evaluation, and content moderation.
  10. Agent did · +11h 48m
    Closed with low-level performance questions: Flash Attention, custom CUDA kernels, GPU-architecture tuning, multi-GPU tensor parallelism, and attention optimization for long sequences.
  11. Outcome · +12h 59m
    Session ended after 614 turns of pure Q&A — no files were opened or edited; every turn was an advisory answer, including several the assistant flagged as unrelated to the actual codebase.
VERDICT

I'd start by telling the assistant up front which of these were meant to touch the actual repo versus general ML knowledge, since it kept flagging that ambiguity itself. Nothing here needed a correction because no code was ever written or changed — it was pure consultation, so the 'fix' that would have helped most next time is scoping the session to a narrower goal rather than firing off hundreds of loosely related questions.

I tried this · 0

Discussion · 0

How Builds work