ikhlasulakmalh/multilingual-hivemind-responses

Multilingual Artificial Hivemind: model responses

Responses to the Infinity-Chat100 prompts in each study language, from the multilingual-artificial-hivemind project. Sampling follows the Artificial Hivemind paper's main setting (temperature 1.0, top-p 0.9, at most 2048 new tokens, reasoning turned off where the model allows it) with 20 samples per prompt instead of the paper's 50. Models generated with 50 are cut to their first 20 here, so every model has the same number.

Models: gemma-4-31b, gemma-4-e4b, ministral-3-8b, mistral-small-4, muse-glimmer-30b, qwen3.5-27b, qwen3.5-9b

Languages: ar-EG, en, es, fa, id, jv, ka, kk, ru, su, th, vi, yo

How to load

The dataset is private: log in first with hf auth login (or set HF_TOKEN).

One language, every model, with 🤗 Datasets. The config names are the language codes above; all holds everything.

from datasets import load_dataset

ds = load_dataset("ikhlasulakmalh/multilingual-hivemind-responses", "id", split="train")
df = ds.to_pandas()
df[df.model == "gemma-4-31b"].head()

One model and language straight into pandas, without Datasets:

import pandas as pd

df = pd.read_parquet("hf://datasets/ikhlasulakmalh/multilingual-hivemind-responses/responses/gemma-4-31b/id.parquet")

The 20 samples for one prompt, only those GlotLID agrees are in Indonesian:

p = df[(df.prompt_id == "12") & (df.on_language != False)]
for text in p.response:
    print(text[:200])
    print("---")

on_language is null where language ID has not run yet, so != False keeps those rows.

Embeddings, lined up with the responses:

import numpy as np
from huggingface_hub import hf_hub_download

path = hf_hub_download("ikhlasulakmalh/multilingual-hivemind-responses", "embeddings/bge-m3/gemma-4-31b/id.npy", repo_type="dataset")
emb = np.load(path).astype(np.float32)   # (prompts, 20, dim)
prompt_ids = list(df.prompt_id.unique())  # same order as emb's first axis

# the paper's hivemind measure for one prompt: mean cosine between its samples
e = emb[prompt_ids.index("12")]
sim = e @ e.T
print((sim.sum() - np.trace(sim)) / (len(e) * (len(e) - 1)))

emb[i, s] is sample s of the i-th prompt in the Parquet file.

Download everything (about 2 GB) to a local folder:

hf download ikhlasulakmalh/multilingual-hivemind-responses --repo-type dataset --local-dir multilingual-hivemind

Responses

responses/<model>/<lang>.parquet, one row per sample. The all config holds every model and language; each language also has its own config.

column
model, hf_id	model key and Hugging Face repo
lang, prompt_id	language code and Infinity-Chat100 prompt id
variant	source (English), translated or localized
prompt	the prompt as given to the model
sample	sample index, 0-19
response	final answer, reasoning stripped
raw	untouched decode
finish_reason, num_tokens, truncated	truncated is true at the 2048-token cap
glotlid_label, glotlid_prob	GlotLID language of the response; null where not yet labelled
on_language	false only when GlotLID gives another language with p >= 0.5 on at least 20 characters

Embeddings

embeddings/<embedder>/<model>/<lang>.npy: float16 array of shape (prompts, 20, dim), L2-normalised, prompts in the same order as the Parquet file. Embedders: bge-m3 (main), octen-emb-4b (robustness check) and gemini-emb-2-instr (Gemini Embedding 2 with the task "sentence similarity" in the text, best in our FLORES-200 check; gemma-4-e4b and qwen3.5-27b only).

Scores

scores/<embedder>.jsonl (per model, language and prompt: within, within_onlang, baseline) and agreement/<embedder>.jsonl (intra- and inter-model agreement), as described in the project README.