latch-eval-tools 0.1.16.dev1

Shared eval tools for single-cell bench, spatial bench, and future biology benchmarks.

latch-eval-tools

Shared eval tools for single-cell bench, spatial bench, and future biology benchmarks.

Installation

pip install latch-eval-tools

Components

Types

from latch_eval_tools import Eval, EvalResult

eval_case = Eval(
    id="test_001",
    task="Count cells in the dataset",
    data_node="latch:///data/sample.h5ad",
    grader={"type": "numeric_tolerance", "config": {...}}
)

Graders

Available graders: numeric_tolerance, label_set_jaccard, distribution_comparison, marker_gene_precision_recall, marker_gene_separation, spatial_adjacency, multiple_choice

from latch_eval_tools.graders import get_grader, NumericToleranceGrader

grader = get_grader("numeric_tolerance")
result = grader.evaluate(
    agent_answer={"n_cells": 1523},
    config={
        "ground_truth": {"n_cells": 1500},
        "tolerances": {"n_cells": {"type": "relative", "value": 0.05}}
    }
)
print(result.passed)
print(result.reasoning)

Harness

Run evaluations with different agents:

from latch_eval_tools.harness import EvalRunner, run_minisweagent_task

runner = EvalRunner("evals/count_cells.json", cache_name=".scbench")
result = runner.run(agent_function=lambda task, work_dir: 
    run_minisweagent_task(task, work_dir, model_name="anthropic/claude-sonnet-4")
)

def my_agent(task_prompt: str, work_dir: Path) -> dict:
    return {"answer": json.loads((work_dir / "eval_answer.json").read_text())}

runner.run(agent_function=my_agent)

Built-in agents: run_minisweagent_task, run_claudecode_task, run_plotsagent_task

Linter

Validate eval JSON files:

eval-lint evals/my_dataset/
eval-lint evals/ --format json

from latch_eval_tools.linter import lint_eval, lint_directory

result = lint_eval("evals/test.json")
print(result.passed, result.issues)

Eval JSON Schema

{
  "id": "unique_test_id",
  "task": "Task description for the agent",
  "data_node": "latch:///path/to/data.h5ad",
  "grader": {
    "type": "numeric_tolerance",
    "config": {
      "ground_truth": {"field": 42},
      "tolerances": {"field": {"type": "absolute", "value": 1}}
    }
  }
}