hf-explorer

name	HF Explorer
description	Mount and explore Hugging Face Hub repos as local filesystems via hf-mount

HF Explorer

You have tools to mount Hugging Face Hub repos as local filesystems. Once mounted, files are fetched lazily — use your normal file tools (Read, ls, Glob, grep) to explore.

When to Activate

• User mentions a HF repo (e.g. meta-llama/Llama-3.1-8B, openai/whisper-large-v3)
• User needs a model for a task — search, then mount the best candidate
• User asks about model architecture, config, tokenizer, or dataset structure
• User wants to compare models or inspect what's inside a repo

Workflow

1. hf_search(query) — Find repos by keyword, task, or model family. Start here when you don't have a specific repo.
2. hf_mount(repo_id) — Mount the repo locally. Returns the mount path.
3. Explore with native tools — ls, Read, Glob, grep on the mount path. Files load on demand.
4. hf_unmount(repo_id) — Clean up when done.

Examples

Find and inspect a model:

hf_search("code generation python") → pick best match →
hf_mount("bigcode/starcoder2-15b") → mounted at /tmp/hf-mounts/bigcode_starcoder2-15b →
ls /tmp/hf-mounts/bigcode_starcoder2-15b →
Read config.json, tokenizer_config.json

Explore a dataset:

hf_search("sentiment", type="dataset") →
hf_mount("stanfordnlp/imdb", type="dataset") →
ls, read sample files

Key Files in Model Repos

• config.json — architecture, hidden size, layers, vocab
• tokenizer_config.json — tokenizer type, special tokens
• generation_config.json — default generation params
• README.md — model card, benchmarks, usage examples

Tips

• Files are lazy — reading config.json doesn't download model weights
• Sort search by trending for what's hot, downloads for battle-tested
• Filter searches: hf_search(query, filter="task:text-generation")
• Private repos need HF_TOKEN env var
• Mount stays active until you unmount or the session ends