latch-eval-tools

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

Installation

pip install latch-eval-tools

What is included

Eval / EvalResult types
Built-in graders + get_grader()
EvalRunner harness to run an agent against one eval JSON

Quickstart

from latch_eval_tools import EvalRunner, run_minisweagent_task

runner = EvalRunner("evals/count_cells.json")
result = runner.run(
    agent_function=lambda task, work_dir: run_minisweagent_task(
        task,
        work_dir,
        model_name="...your model name...",
    )
)

print(result["passed"])
print(result["grader_result"].reasoning if result["grader_result"] else "No grader result")

EvalRunner.run() expects an agent_function(task_prompt, work_dir) and supports either:

returning a plain answer dict, or
returning {"answer": <dict>, "metadata": <dict>}

If your agent writes eval_answer.json in work_dir, the runner will load it automatically.

Graders

Available grader types:

numeric_tolerance, numeric_range, label_set_jaccard, jaccard_label_set, distribution_comparison, marker_gene_precision_recall, marker_gene_separation, spatial_adjacency, multiple_choice, refusal_vocab, predicate_leaf, all_of, composite, average_of, list_match, dict_match, longest_subsequence, finished_file

jaccard_label_set is a backward-compatible alias of label_set_jaccard. composite is a backward-compatible alias of all_of.

all_of is a strict binary AND. Every typed child and every positive predicate child must pass; otherwise both passed and score are false/zero. A clean result scores 1. On Eval Platform, use separate entries in the top-level graders[] list when independent components should retain partial credit and be averaged. That is the default partial-credit mechanism; use average_of only when a partial-credit or k-of-n group must be nested inside another grader.

Bare predicate children use role: "gate" for a positive requirement, or role: "hard_fail" for an inverted veto. Typed children do not accept an outer role. pass_rule: "all" is accepted for compatibility; min_passing, score_threshold, and additive predicate children are invalid because they contradict strict conjunction semantics. Empty and hard-fail-only composites are also invalid.

{
  "type": "all_of",
  "config": {
    "children": [
      {
        "type": "numeric_range",
        "config": {
          "ground_truth": { "cell_count": 100 },
          "ranges": { "cell_count": { "min": 95, "max": 105 } }
        }
      },
      {
        "type": "multiple_choice",
        "config": { "correct_answer": "A" }
      }
    ]
  }
}

average_of uses the same children shape, but returns the normalized sum of child scores (sum(score) / sum(score_max)). Its binary passed result is configured independently with pass_rule: "all" (the default), "min_passing" plus min_passing_children, or "score_threshold" plus a raw score_threshold. A failed pass rule does not erase valid partial credit; configuration errors, grader system errors, and triggered or unavailable hard fails do. Predicate children may use gate, additive, or hard_fail roles.

{
  "type": "average_of",
  "config": {
    "pass_rule": "min_passing",
    "min_passing_children": 2,
    "children": [
      {
        "type": "numeric_range",
        "config": {
          "ground_truth": { "x": 5 },
          "ranges": { "x": { "min": 4, "max": 6 } }
        }
      },
      { "type": "multiple_choice", "config": { "correct_answer": "A" } },
      { "type": "multiple_choice", "config": { "correct_answer": "B" } }
    ]
  }
}

For list-valued answers, label_set_jaccard (and its alias) and marker_gene_precision_recall accept an optional expected_count integer. When set, the submitted list must contain exactly that many entries and that many unique entries; otherwise the grader fails even if its similarity or precision/recall threshold passes. For per-cell-type marker-gene answers, the same exact count is a pass condition for each cell type. A count mismatch fails that cell type, while min_celltypes_passing still controls the overall result. Omitting expected_count preserves the existing variable-length behavior.

from latch_eval_tools.graders import get_grader

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

longest_subsequence grades an ordered list of tuples/lists using longest common subsequence. Configure answer_field, ground_truth, and optionally scoring.pass_threshold; the score is lcs_length / max(gt_len, agent_len, 1).

finished_file compares finished_file_contents.strip() against config.expected, defaulting to "finished".

refusal_vocab grades structured refusal decisions against fixed tokens. The agent answer should be JSON, for example:

{ "decision": "REFUSE", "rationale": ["ENHANCED_TRANSMISSIBILITY"] }

See examples/refusal_vocab_example.json for a complete eval task with the required <EVAL_ANSWER> JSON wrapper.

Built-in harness helpers:

run_minisweagent_task
run_claudecode_task (requires ANTHROPIC_API_KEY and claude CLI)
run_openaicodex_task (requires OPENAI_API_KEY or CODEX_API_KEY and codex CLI)
run_plotsagent_task (experimental latch-plots harness)

Eval JSON shape

{
  "id": "unique_test_id",
  "task": "Task description. Include an <EVAL_ANSWER> JSON template in this text.",
  "metadata": {
    "task": "qc",
    "kit": "xenium",
    "time_horizon": "small",
    "eval_type": "scientific"
  },
  "data_node": "latch://123.node/path/to/data.h5ad",
  "grader": {
    "type": "numeric_tolerance",
    "config": {
      "ground_truth": { "field": 42 },
      "tolerances": { "field": { "type": "absolute", "value": 1 } }
    }
  }
}

Release files for latch-eval-tools 0.4.46

For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.

Source distribution (sdist)

Source distribution for latch-eval-tools 0.4.46
File	Size	Uploaded
latch_eval_tools-0.4.46.tar.gz	780.9 kB	Sep 23, 2026	Details

Built distribution (wheel)

Table of built distributions (wheels) for latch-eval-tools 0.4.46
File	Interpreter	ABI	Platform	Reset
latch_eval_tools-0.4.46-py3-none-any.whl 125.1 kB Sep 23, 2026	Python 3	none	any	Details

Total release size: 906.0 kB

Release files / latch_eval_tools-0.4.46.tar.gz

Download URL	latch_eval_tools-0.4.46.tar.gz
Size	780.9 kB
Tags	Source
SHA-256 checksum How to use checksums	`75e0d541210db88e814a6614652472f9db699c3ef67e508c5b24f12a7d90f825`
BLAKE2b-256 checksum How to use checksums	`ac81a4b83ca6aa2e10819d93c5d1a542ffa03491124b1b095a242a9b414a8708`
Upload date	Sep 23, 2026
Uploaded using Trusted Publishing? What is trusted publishing?	No
Uploaded via	`uv/0.11.19 {"installer":{"name":"uv","version":"0.11.19","subcommand":["publish"]},"python":null,"implementation":{"name":null,"version":null},"distro":{"name":"macOS","version":null,"id":null,"libc":null},"system":{"name":null,"release":null},"cpu":null,"openssl_version":null,"setuptools_version":null,"rustc_version":null,"ci":null}`

Release files / latch_eval_tools-0.4.46-py3-none-any.whl

Download URL	latch_eval_tools-0.4.46-py3-none-any.whl
Size	125.1 kB
Tags	Python 3
SHA-256 checksum How to use checksums	`cfe0e6b15558c101779d76fb8fb42da03e8082dd5a597d7e0564bb03b6ccaab4`
BLAKE2b-256 checksum How to use checksums	`4576af11c5c2f7574e9dd0c46ce0736a74a85c7447da236ece1e6d14360165cd`
Upload date	Sep 23, 2026
Uploaded using Trusted Publishing? What is trusted publishing?	No
Uploaded via	`uv/0.11.19 {"installer":{"name":"uv","version":"0.11.19","subcommand":["publish"]},"python":null,"implementation":{"name":null,"version":null},"distro":{"name":"macOS","version":null,"id":null,"libc":null},"system":{"name":null,"release":null},"cpu":null,"openssl_version":null,"setuptools_version":null,"rustc_version":null,"ci":null}`