v1.4.0.rc0
📦 huggingface-hubView on GitHub →
✨ 10 features🐛 4 fixes🔧 16 symbols
Summary
This release introduces significant enhancements to the CLI, including a new skill installation command (`hf skills add`), improved help output, and new commands for managing papers and collections. It also adds a decentralized evaluation results module to the Hub with corresponding Python helpers.
Migration Steps
- If using AI coding assistants, run `hf skills add --<agent_name>` (e.g., `hf skills add --claude`) to enable new Hub interaction capabilities.
- If uploading evaluation results programmatically, update serialization/parsing logic to use `EvalResultEntry`, `eval_result_entries_to_yaml()`, and `parse_eval_result_entries()`.
- When running multi-GPU training jobs via CLI, use the `--with torch` or `--with accelerate` flags followed by the appropriate launcher command (e.g., `hf jobs uv run --with torch -- torchrun train.py`).
✨ New Features
- Introduced the `hf skills add` CLI command to install the `hf-cli` skill for AI coding assistants (Claude Code, Codex, OpenCode), enabling capabilities like Hub searching, model downloading, and Job management.
- The `hf-cli` skill is installed to a central `.agents/skills/hf-cli/` directory with relative symlinks created for agent-specific directories to ensure a single source of truth.
- Improved CLI help output organization, grouping commands into 'Main commands' and 'Help commands', and adding 'Examples' and 'Learn more' sections.
- Implemented a decentralized system for tracking model evaluation results via `.eval_results/*.yaml` files in model repos, which automatically populate leaderboards on benchmark datasets.
- Added Python helpers in `huggingface_hub` to manage evaluation results: `EvalResultEntry` dataclass, `eval_result_entries_to_yaml()`, and `parse_eval_result_entries()`.
- Added `hf papers ls` command to list daily papers on the Hub, supporting filtering by date and sorting by trending or publication date.
- Added `hf collections` commands (`ls`, `create`, `add-item`, `info`, `delete`) for managing collections via the CLI.
- Multi-GPU training commands are now supported in `hf jobs uv run` using `--with torch` for `torchrun` or `--with accelerate` for `accelerate launch`.
- The `hf jobs uv run` command now supports passing local configuration files alongside scripts.
- Introduced the `hf jobs hardware` command to list available hardware options for jobs, including specifications and pricing.
🐛 Bug Fixes
- CLI output format option added for `ls` commands.
- CLI implemented dynamic table columns based on the `--expand` field.
- Centralized error handling added to the CLI.
- Improved exception handling scope within the CLI.