hive-cascade / README.md
SkyIsNotGreen's picture
The paper lives in the Hive repo at cascade/CASCADE-PAPER.md
866b932 verified
|
Raw History Blame Contribute Delete
4.46 kB
metadata
license: mit
task_categories:
  - text-generation
  - question-answering
language:
  - en
tags:
  - evaluation
  - llm
  - local-first
  - cascade
  - gold-free
  - verification
pretty_name: Hive Cascade task sets
size_categories:
  - n<1K

license: mit task_categories: - text-generation - question-answering language: - en tags: - evaluation - llm - local-first - cascade - gold-free - verification pretty_name: Hive Cascade task sets size_categories: - n<1K

Hive Cascade task sets

Two evaluation sets used to measure the Hive Cascade: a 100-task bench and a 664-task community set. Each task is a prompt, a checker and the checker's expected value. The runtime protocol never sees the checker; the evaluation pass reads it only after the answers are fixed (the gold-free contract).

The paper is not this card. It lives in the repository: cascade/CASCADE-PAPER.md. The visual companion is the static Space SkyIsNotGreen/hive-cascade-paper.

Contents

the file n the sources
tasks-bench.json 100 gsm8k 40, MATH 25, HumanEval 20, TruthfulQA 15
tasks-community.json 664 gsm8k 300, MATH 200, HumanEval 164

Splits: bench 70 train / 30 test; community 465 train / 199 test. Buckets: reasoning (gsm8k + MATH), code (HumanEval), qa (TruthfulQA, bench only).

The two files under results/ are the runner's reports from the 2026-10-18 sampled runs (the 35B/27B cascade), included as provenance for the paper's numbers.

The task format

{
  "id": "gsm8k-0005",
  "benchmark": "gsm8k",            // gsm8k | math | humaneval | truthfulqa
  "source": "openai/gsm8k",        // the source dataset and, for MATH, the subject
  "bucket": "reasoning",           // reasoning | code | qa
  "prompt": "Rory is retrieving ...",
  "checker": {"type": "number", "expect": 10.0},
  "meta": {"source_split": "test"},
  "split": "test"
}

The checker is one of four kinds:

checker.type what it does
number compares the final numeric answer with expect
boxed parses the last \\boxed{...} and compares it
code runs the candidate against the HumanEval harness (prelude + tests, entry_point)
choice compares the multiple-choice letter with expect

The checkers carry the gold on purpose. They are the evaluation instrument, not the runtime gate; the cascade's runtime uses verifiers that never read them.

Loading

import json
from huggingface_hub import hf_hub_download

path = hf_hub_download("SkyIsNotGreen/hive-cascade", "tasks-bench.json",
                       repo_type="dataset")
tasks = json.load(open(path))["tasks"]
train = [t for t in tasks if t["split"] == "train"]

Sources and attribution

The task text is redistributed from:

The packaging and the metadata are MIT; follow each source's licence for the task text.

Provenance

  • built 2026-10-02 (bench) and 2026-10-07 (community), seed 0;
  • the bench is a stratified 100-task cut of the four sources; the community set is the larger gsm8k/MATH/HumanEval pool;
  • the paper's runs used these files unchanged (their sha256 is in the runner reports).

Links

the paper cascade/CASCADE-PAPER.md
the visual companion hive-cascade-paper
the code and the full results log github.com/sky-is-green/hive
the memory system SkyIsNotGreen/splinter-memory
the encoder (research checkpoint) SkyIsNotGreen/splinter-encoder