Survey ML tooling, models, and infra options for an AI dashboard repo
Session from hf-coding-tools-traces-all on Hugging Face by davidkling · CC-BY-4.0 · condensed by Coders Talk
Click a segment to jump to that moment
35msession
0interventions
GOAL
The developer wanted broad guidance on datasets, models, fine-tuning, deployment, and tooling choices for AI/ML work, apparently while exploring what to build next in a Next.js-based dashboard repo.
- +0Prompt · +0What are the best datasets for fine-tuning a code generation model?
- +0Agent did · +0Answered a long run of general ML questions on datasets, instruction tuning, chatbots, evaluation, and best current LLMs/embedding models with generic knowledge-based advice, no repo access used.
- +4mFail · +4mRepeatedly hit 'sandbox restrictions' when asked to search the repo for MMLU/HumanEval, CI configs, safetensors references, Llama 3 mentions, or checkpoint code, and had to ask the developer to paste files instead.
- +6mPrompt · +6mHow do I add AI-powered text summarization to my web app?
- +6mAgent did · +6mSketched high-level plans for adding summarization, RAG chatbots, LLM API integration, Gradio demos, and voice-to-text into the (Next.js-variant) app, all as advice with no code written or files touched.
- +7mPrompt · +7mWhat's the best enterprise platform for deploying custom LLMs?
- +7mAgent did · +7mCovered enterprise deployment platforms, SOC2 compliance, quantization, LoRA/QLoRA fine-tuning, distributed training, and GPU serving strategies, again purely as generic guidance.
- +22mPrompt · +22mHow do I train a custom tokenizer for my domain-specific data?
- +22mAgent did · +22mContinued through tokenizers, vision models, embeddings, AI agents, robotics, experiment tracking, safety/alignment, model format conversion, browser/on-device inference, and GPU kernel optimization — all conceptual Q&A, no repo changes at any point.
- +35mOutcome · +35mSession ended after 526 turns of pure Q&A; no files were edited, no code was written, and no PR or test came out of it — the repo was untouched throughout.
VERDICT
I'd start by telling the agent up front which of these ~130 questions actually matters for the repo I'm working in, instead of firing off a long, unrelated stream of general ML trivia. The sandbox restrictions kept blocking any real repo inspection, and I never corrected that by pasting in the files it asked for, so the whole session stayed advice-only with zero code touched.