Skip to content
PitestsData & ML1 file · +22 −6

Find aerial detection/segmentation datasets on HF via CLI

Session from hf-find-traces on Hugging Face by merve · Apache-2.0 · condensed by Coders Talk

3msession
1interventions
GOAL

Search Hugging Face Hub for aerial object detection and segmentation datasets using the hf CLI, and ensure future sessions use the better semantic search tool.

  1. Prompt · +0
    Find aerial object detection segmentation datasets on HF
  2. Agent did · +0m
    Tried several hf CLI subcommands (dataset-search, datasets search, datasets list --search) with various queries like xview, dota, visdrone, hrsc - most returned no results or errors
  3. Agent did · +0m
    Continued keyword search with hf datasets list --search 'aerial', got some results and summarized a table of aerial detection datasets
  4. Prompt · +0m
    Asked if hf-find is installed with pi
  5. Prompt · +1m
    Pointed out hf-find is an hf extension and asked to run hf find --help
  6. Agent did · +1m
    Ran hf find --help, discovered hf-find extension exists, used hf find datasets semantic search, got much better matching results (isaid, dota-v1.5, SkyScenes, etc.)
  7. Intervention · +2m
    Asked why the agent didn't use hf-find from the start since it's already installed
    Why: It was already installed - the agent should have checked for better tools before defaulting to weak keyword search.
  8. Prompt · +2m
    Asked what to do so pi sees hf-find next time
  9. Agent did · +2m
    Read the hf-cli SKILL.md, found it lacked hf-find extension docs, and edited the skill file to document hf find commands for future sessions
  10. Outcome · +3m
    Updated SKILL.md with hf find semantic search commands (+22 −6 lines) so future sessions default to the better search tool
VERDICT

Next time I'd just ask upfront whether there's a better search tool installed instead of letting the agent grind through keyword search failures. Pointing out that hf-find was already installed was what actually fixed things - once I nudged it to check, the semantic search immediately gave better results. Updating the skill doc afterward should prevent this same detour from happening again.

I tried this · 0

Discussion · 0

How Builds work