feat: HuggingFace dataset import command

[christso]

Objective

Add an agentv import huggingface CLI command to import datasets from HuggingFace Hub directly into agentv eval format.

Motivation

SWE-bench and similar benchmarks publish their datasets on HuggingFace (e.g., SWE-bench/SWE-bench_Verified, princeton-nlp/SWE-bench). Currently users must manually convert these to agentv YAML/JSONL format. A built-in importer would remove this friction and make agentv a first-class tool for benchmark evaluation.

Design

This should be a CLI wrapper (not a core feature), following the "Lightweight Core, Plugin Extensibility" principle.

Proposed interface

agentv import huggingface --repo SWE-bench/SWE-bench_Verified --split test --limit 10 --output evals/swebench/

Mapping (SWE-bench → agentv)

SWE-bench Field	AgentV Field
instance_id	tests[].id
problem_statement	tests[].input[0].content
repo + base_commit	workspace.docker.image (convention-based)
FAIL_TO_PASS	tests[].assertions[].command (code-grader)
difficulty	tests[].metadata.difficulty

Implementation approach

• Python script using uv run (per repo convention for Python scripts)
• Uses datasets library to load from HuggingFace
• Outputs .EVAL.yaml files with Docker workspace configs
• Template-based: support different dataset schemas via mapping configs

Acceptance Criteria

• agentv import huggingface --repo <name> produces valid eval YAML files
• Works with SWE-bench Verified as the primary test case
• Supports --limit, --split, --output flags
• Generated evals pass agentv validate
• Documentation updated

Non-goals

• Not adding HuggingFace as a core dependency
• Not supporting all possible HuggingFace dataset formats (start with SWE-bench)