𒁵𒄩 (munku)

Decision Index 0.3 public score: 71.10 (self-run).

• A fine-tune of google/gemma-4-31B-it that builds on plantwaterer/eshbar-lulilla, with its own last-layer head.
• Engine: munku:Engine, in-process, from this repository: eshbar-lulilla's engine as first released (without its later draft head), with its files renamed (names only). Engine, head, store and executor files are pinned by MD5 in PINS.json and checked at load.
• Kit: apolinario/decision-index @ 9eb2dbe2a358004c8782c66e40a83ac07b953fec.

Usage

REPO=$(python -c "from huggingface_hub import snapshot_download; print(snapshot_download('plantwaterer/munku'))")
PYTHONPATH="$REPO" python -m decision_index run --engine munku:Engine --option model="$REPO" --option route=off --compact --out runs/munku

Inference

• Precision, attention, averaging, head: bf16; packed document attention (full-attention layers bidirectional within a prompt, sliding-window layers causal); for requests under a length limit, the answer is averaged over 4 option orders. Ordinary choices: the letter logits' softmax at one fitted temperature T = 1.2786 (no posterior). Yes / no decisions: a Laplace posterior on binary (yes / no) decisions, integrated.
• Knowledge store and executed answers: off. The engine carries eshbar-lulilla's knowledge store and executed-answer route (a numeric or yes / no question the first pass is unsure of goes to a bundled store of country-level statistics, then to a short Python function the model writes, run by a bundled exact-arithmetic executor). This model runs with it switched off (--option route=off), so every answer is the first pass's. The public run was made with it on; it changed no answer (0 of 3,490 routed questions produced a usable program, and the store answered none).
• Hardware: one GPU with at least 75 GB (bf16 weights 62 GB); fits one RTX PRO 6000 (96 GB). The executor's kernel ships prebuilt (sm_75 to sm_120).

Latency (one RTX PRO 6000 Blackwell, 1,000-request sample)

Median 291 ms, mean 652 ms, p80 871 ms, p95 1,848 ms on the board's 1,000-row public sample, one request at a time, with --option route=off. On a second sample of 1,000 held-out questions: median 192 ms, mean 267 ms, p80 292 ms. Model load and warm-up (up to about three minutes) are not included.

Third-party components

• munku/xr_runtime/, the executor: derived from LOOM by Mehmet Kerem Turkcan and contributors (https://github.com/mkturkcan/Loom), Apache License 2.0, modified. Its LICENSE and NOTICE are in that directory.
• store_table.json, the knowledge store's table: licensed CC BY-SA 4.0. It contains data from Our World in Data (CC BY 4.0), compiled from the producers listed in DATA_NOTICE, including V-Dem (CC BY-SA 4.0). Values are rounded to three significant figures and re-indexed, so the table is a modified version of the original data. Full attribution: DATA_NOTICE.

Training data

munku builds on plantwaterer/eshbar-lulilla, itself continued from plantwaterer/liquamen; the training data, licences and attributions listed on those two cards carry over. The training added for this model used 384,663 decision questions: the eshbar-lulilla mixture (which carries liquamen's sources), the additions below, and synthetic questions generated by the author. Row counts below are of those 384,663 questions.

Some sources carry non-commercial or research-only terms, and some state no licence:

• Non-commercial: ANLI (Nie et al., 2020; CC BY-NC 4.0; 702 rows); AmericasNLI (CC BY-SA 4.0), a translation of XNLI (Conneau et al., 2018; CC BY-NC 4.0; 11 rows); x-stance (Vamvas & Sennrich, 2020; CC BY-NC 4.0; 400 rows); the Amazon Multilingual Counterfactual Dataset (O'Neill et al., 2021; CC BY-NC 4.0; 400 rows); and, within DISRPT, the Potsdam Commentary Corpus (CC BY-NC-SA; 160 rows), the Russian RST Treebank 1.0 (CC BY-NC-SA 4.0; 105 rows) and the Persian RST corpus (CC BY-NC; 3 rows). RAGTruth's annotations are MIT, but its question-answering passages come from MS MARCO, whose terms limit use to non-commercial research (813 rows).
• Research use: CSTNews (Cardoso et al., 2011; 160 rows), "available for download and use for research purposes".
• States no licence: ACOS (Restaurant-ACOS and Laptop-ACOS, Cai et al., 2021; 1,617 rows), MultiRC (Khashabi et al., 2018; 555 rows), NLI4CT (Jullien et al., 2023; 702 rows), VAST (Allaway & McKeown, 2020; 702 rows) and SciDTB (Yang & Li, 2018; 1 row). No licence in the DISRPT distribution: the RST Basque TreeBank, the Dutch Discourse Treebank and the RST Spanish Treebank (160 rows each).

Added here (the mathematics sources and the phishing-email question wording were also in eshbar-lulilla's training data, though not named on its card):

• Mathematics: the DeepMind Mathematics Dataset (Saxton et al., 2019), problems generated afresh with its own generator (Apache-2.0); AQuA-RAT (Ling et al., 2017; Apache-2.0) and MATH (Hendrycks et al., 2021; MIT) train splits.
• Medicine and logic: MedQA (Jin et al., 2021; MIT) and FOLIO (Han et al., 2022; CC BY-SA 4.0 in the authors' GitHub repository, MIT on their Hugging Face page) train splits.
• Science questions: SuperGPQA (M-A-P Team, 2025; ODC-BY 1.0), released as a single split. Its answer keys were checked with Qwen3-32B (Apache-2.0).
• Tools: Toucan-1.5M (Xu et al., 2025; Apache-2.0): MCP tool schemas with questions and trajectories written by open-weight models, recast as tool-decision questions.
• Fact checking, clinical evidence, intents and products: EX-FEVER (Ma et al., 2024; MIT), with English Wikipedia evidence (CC BY-SA); Evidence Inference 2.0 (DeYoung et al., 2020; MIT), annotator evidence spans only; MTOP (Li et al., 2021; CC BY-SA 4.0); MultiDoGO finance intents (Peskov et al., 2019; CDLA-Permissive-1.0); WANDS (Chen et al., 2022; MIT), labels mapped to the ESCI classes. Corr2Cause (already listed) contributed further train-split items.
• Teacher targets: soft targets from MiMo-V2.6-Pro-RL (Xiaomi; MIT) on 5,354 questions (PubMedQA artificial-label items and one synthetic family).
• Synthetic questions generated by the author (no third-party data), several in the request formats of the Decision Index kit (MIT); the phishing-email questions use the question wording of jev-phishing-bench (MIT). Task formats follow BIG-Bench Hard, MuSR, CLadder, CRUXEval and SATA-Bench; no items from those benchmarks are used.
• Further NLP datasets (train splits, a few hundred questions each), each under its own licence as stated by its authors; where the copy used on the Hub is tagged differently, the authors' statement is given:

dataset (Hugging Face)	licence
amydeng2000/CREAK	MIT
hitachi-nlp/FLD.v2	Apache-2.0 (code), CC BY 4.0 (corpus)
sadat2307/MSciNLI	CC BY-SA 4.0
tasksource/scinli	CC BY-SA 4.0
neuralchemy/Prompt-injection-dataset	Apache-2.0
hendzh/PromptShield	Apache-2.0
S-Labs/prompt-injection-dataset	MIT
deepset/prompt-injections	Apache-2.0
tasksource/SpaRTUN	MIT
mteb/amazon_counterfactual	CC BY-NC 4.0 (authors' LICENSE file)
allenai/art	Apache-2.0
tasksource/avicenna	Apache-2.0
AshtonIsNotHere/biosift-nli	MIT
tasksource/commonsense_qa_2.0	CC BY 4.0
pietrolesci/conj_nli	MIT
tasksource/counterfactually-augmented-snli	Apache-2.0
tasksource/cycic_classification, cycic_multiplechoice	Apache-2.0
fancyzhx/dbpedia_14	CC BY-SA 3.0
AndyChiang/dgen	MIT
multilingual-discourse-hub/disrpt	per corpus: Potsdam Commentary Corpus CC BY-NC-SA, Russian RST Treebank CC BY-NC-SA 4.0, Persian RST corpus CC BY-NC, GCDT Apache-2.0; CSTNews "available for download and use for research purposes"; SciDTB states no licence; the RST Basque TreeBank, the Dutch Discourse Treebank and the RST Spanish Treebank: no licence in the DISRPT distribution
neurae/dnd_style_intents	Apache-2.0
YufeiHFUT/DocRED_origin	MIT (thunlp/DocRED)
hendrycks/ethics	MIT
mwong/fever-evidence-related	CC BY-SA 3.0, GPL-3.0
tasksource/few_rel	MIT
nightingal3/fig-qa	MIT
tasksource/fool-me-twice	Apache-2.0 (Wikipedia evidence CC BY-SA)
WhereIsAI/github-issue-similarity	MIT
tasksource/hans	MIT
EleutherAI/headqa	MIT (aghie/head-qa)
IlyaGusev/headline_cause	CC0 1.0
marcov/health_fact_promptsource	MIT (PUBHEALTH)
tasksource/idioms-nli	MIT
coastalcph/lex_glue	CC BY 4.0
tasksource/math_qa	Apache-2.0 (allenai/math_qa)
mneb/maven	MIT
mteb/AfriSentiClassification	CC BY 4.0
mhardalov/exams	CC BY-SA 4.0
klue/klue	CC BY-SA 4.0
community-datasets/offenseval_dravidian	CC BY 4.0
google-research-datasets/paws, paws-x	Google terms: "may be freely used for any purpose", acknowledgement of Google LLC appreciated
tasksource/x-fact	MIT
michiel/xstance	CC BY-NC 4.0 (ZurichNLP/xstance; data (c) smartvote.ch)
tasksource/nlgraph	MIT
tasksource/nli-veridicality-transitivity	CC BY-SA 4.0
pietrolesci/nli_fever	MIT
tasksource/numer_sense	MIT
Korea-MES/open_question_type	CC BY 4.0 (launch/open_question_type)
allenai/openbookqa	Apache-2.0
tasksource/patent-phrase-similarity	CC BY 4.0
community-datasets/proto_qa	CC BY 4.0
allenai/qasc	CC BY 4.0
allenai/quartz	CC BY 4.0
tasksource/robustLR	MIT
strombergnlp/rumoureval_2019	CC BY 4.0
tasksource/scicite	Apache-2.0
allenai/scitail	Apache-2.0
pacovaldez/stackoverflow-questions	Apache-2.0 (dataset page; the text is Stack Overflow content)
tasksource/tracie	Apache-2.0
tasksource/twentyquestions	Apache-2.0
tals/vitaminc	CC BY-SA 3.0
tasksource/wice	ODC-BY 1.0 (annotations); Wikipedia text CC BY-SA
MoE-UNC/wikihop	CC BY-SA 3.0 (QAngaroo)
sileod/wikimedqa	Apache-2.0
Salesforce/wikisql	BSD-3-Clause
allenai/wiqa	Apache-2.0

Benchmarks with a train split contributed train-split records only. SuperGPQA has no separate train split; its use is described above.

Capacity limits

• Context: max_len 32,768 tokens; nothing truncated.