Survey open questions on ML datasets, models, and deployment
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 a broad rundown of current best practices and options across ML dataset sourcing, model selection, fine-tuning, deployment, and agent-building topics.
- +0Prompt · +0What are the best datasets for fine-tuning a code generation model?
- +0Agent did · +0Answered a long run of questions on dataset discovery, versioning, hosting, licensing, and streaming (Hugging Face, DVC, S3, parquet, etc.)
- +3mAgent did · +3mCovered model selection and evaluation topics: benchmarks, eval frameworks, cost tradeoffs, quantization, and picking models for specific tasks (NER, TTS, classification, translation)
- +14mAgent did · +14mWalked through deployment and serving concerns: inference servers, quantization formats, Kubernetes/GPU scheduling, serverless endpoints, and fine-tuning methods (LoRA, QLoRA, DPO)
- +21mAgent did · +21mDiscussed building agents and demos: multi-tool agents, RAG chatbots, robot learning, embeddings, tokenizers, and experiment tracking
- +28mAgent did · +28mClosed with safety/alignment practices, model conversion formats, browser-side inference, and low-level GPU optimization (Flash Attention, CUDA kernels, tensor parallelism)
- +35mOutcome · +35mSession ended after 526 turns of Q&A with no code written or files changed — purely a reference dump across datasets, models, deployment, and optimization topics
VERDICT
I'd start by narrowing the scope up front instead of letting it sprawl into 500+ turns of loosely related questions — I never actually asked it to build or change anything, so there was nothing to correct or verify. Next time I'd pick one concrete task (e.g. 'help me fine-tune this specific model on this dataset') and iterate on that, rather than treating it like a search engine for best-practices lists.