logitshq/Logits-Sigma-1-Duo

Logits Sigma-1 Duo

Logits Sigma-1 Duo answers Decision Index requests with two decision models on one GPU. It averages their per-option probabilities 50 / 50 and answers with the argmax; it generates nothing.

Member	Directory	What it is	Weights here
A	a/	PelaAI/KnowLine-31B-Gen1 @ 8f052a81 with Logits' LoRA (v7 warm (final adapter, step 600)) merged at scale 0.5. Read with Logits' prompt layout (v5, option codes, shared-prefix packing)	FP8 block-128 (e4m3, per-128×128 scales), 33.3 GB
B	d3-ultra/	vllm-sr/d3-ultra @ 37fee019 (the revision the board scored) with its own runtime and readout (d3_*.py, unchanged)	FP8 block-128 converted from the bf16 release, 27.8 GB

Logits trained only the LoRA in member A. Both members are otherwise other teams' work, credited below.

Results

Decision Index 0.3 public run	72.93 (raw 79.54). Complete: 140,178 of 140,178 scoreable requests, 0 unsupported, 0 errors. Results: logits-sigma-1-duo-decision-index (compact: no suite text)
Latency, latency-v1 design, one request at a time, 1× RTX PRO 6000 Blackwell Server Edition	median 293 / mean 859 / p80 659 ms. Peak GPU memory 87,891 MiB
Decision agreement on those 750 rows	99.40% vs the members in bf16 (FP8 bar 96%); 99.43% (on the 689 rows also in the public run) vs the same engine on an H200

Run it for the Decision Index

Docker (the serving we measured; it compiles causal-conv1d for d3-ultra's short convolutions):

hf download logitshq/Logits-Sigma-1-Duo --revision d9641cd862fc0ee4833b2fc52142802ba8181c40 --local-dir duo
docker build -f duo/code/Dockerfile --build-context kit=<decision-index @ 25c6ec7> -t logits-sigma-1-duo duo/code
docker run --gpus all -v $PWD/duo:/models/duo:ro -v $PWD/runs:/runs logits-sigma-1-duo \
    pipeline --compact --out /runs/logits-sigma-1-duo

Or pip, Python 3.12, NVIDIA driver ≥ 580:

uv venv --python 3.12 venv && . venv/bin/activate
uv pip install -r duo/code/requirements.lock ./duo/code
uv pip install --no-deps "decision-index @ git+https://github.com/apolinario/decision-index@25c6ec7"
CAUSAL_CONV1D_FORCE_BUILD=TRUE uv pip install --no-build-isolation causal-conv1d==1.7.0   # needs nvcc 13 + g++
python -m decision_index pipeline --engine logits_sigma_1_duo.engine_pair:Duo --option model=duo --compact \
    --out runs/logits-sigma-1-duo

At load the engine prints which FP8 matmul (cuBLAS block-scaled or the vendored Triton kernel) and which d3-ultra kernels (flash-linear-attention, causal-conv1d, or the PyTorch reference) serve the run. environment.json keeps both under model_source.

How it decides

• Member A (logits_sigma_1_duo.engine:LogitsEngine) uses Gemma 4's chat template with thinking off and one prompt per question. The options are labelled with one-token codes. It reads the codes' logits at the answer position and softmaxes them over the question's codes. A request's questions share the state block, which is read once per request (packs of ≤ 8,192 tokens, forwards of ≤ 16,384).
• Member B (d3_engine:D3Engine, vllm-sr's code) is a code readout with prefix-bidirectional attention over the state, from the d3-ultra repo.
• Mixing: for each choice question, the mean of the two members' option probabilities; for each yes / no question, the mean of p(yes). The weights are fixed at 0.5 / 0.5 and were never tuned per benchmark.
• One size rule: a request longer than 50,000 characters (compact JSON) is answered by member A alone. d3-ultra's bidirectional state attention needs ~108 GiB there. The rule depends only on request size, never on an answer; it touches 0.27% of the latency sample.
• Serving: both members on one GPU, member B on its own CUDA stream. FP8 weights; bf16 activations; float32 readout.
• Never: truncation, option dropping, per-benchmark prompts, labels at inference, or retries after a wrong answer.

Training and selection (member A's LoRA)

Logits trained one LoRA (rank 64, α 128, on every language-model projection) in two stages, both on google/gemma-4-31B-it @ 842da379. It was then merged into KnowLine-31B-Gen1 for member A:

• v6, the same adapter as Logits Sigma-1. It trained on 8× H200 at learning rate 1e-4 for 1,500 steps, about 78% of one epoch over 476,186 rows (480M tokens). The loss is soft cross-entropy over the option codes, plus 1.0 × KL to the frozen base, plus 0.3 × Brier.
• v7 warm (the final adapter, step 600) continued v6's adapter at learning rate 5e-5 with the same loss, for 600 steps: about half of one epoch over 158,750 rows. Those are 52,167 new teacher rows and 106,583 rows replayed from v6's file (about a third of the tokens). The final training loss was 0.379.
• Merge: W' = W_KnowLine + 0.5 · (α / r) · B·A on every adapted projection. It is computed in fp32 and rounded once to bf16, then converted to FP8 block-128.

Data, both stages:

• tasksource-jev-typed-decisions: 411 sources tagged for commercial use (151,808 rows in v6, 34,231 in v7 warm). DATA_SOURCES.md lists each with its licence and its rows in each stage.
• Decision Index benchmark train splits in the suite's request format (ditrain: 38,278 rows in v6, 8,555 in v7 warm). Generator families written for benchmark skills; they copy no benchmark item.
• Our generators over public records (NHTSA, OSHA, FJC, openFDA, CVE / CWE / CISA KEV, USPTO), research datasets and the Public Jira Dataset, plus procedural generators that use no third-party data.
• Teacher rows from Logits' pipeline on its own fleet: 64,862 in v6 and 66,599 in v7 warm, 52,167 of them new. Kimi K3.1 Pro writes the task families and DeepSeek V4.1 Flash writes the states; the labels are both models' votes.
• No text or label in either stage was written by a Claude (Anthropic) or OpenAI model. No output of another decision model is a label.
• Decontamination: test and validation splits of Decision Index source datasets were never read. Every generator and teacher output was matched (13-gram and exact) against the rebuilt public 0.3 suite rows and the evaluated splits, and matches were dropped.

How the Duo was chosen (docs/37 in Logits' research repo):

• Members and weights were picked on Logits' mini-suite, which is a subset of the public 0.3 suite, plus two held-out decision sets. The rules were pre-registered before the GPU runs, and the choice between members follows a paired-bootstrap court.
• So the public index is partly in-sample for the selection, and we say so here.
• Neither member was trained or tuned on any private Decision Index test.

Licence and training-data disclosure

Logits' part is released under Apache-2.0: the LoRA merged into member A, the prompt layout, and the engine and pair code in code/. The upstream weights are Apache-2.0 too (credited below), so the repository is Apache-2.0 throughout.

This release follows decisions.md D18 (Ryan, 2026-11-07: release under Apache-2.0 with disclosure). Here is what member A's LoRA's training data holds:

• Share-alike sources. Both stages include CC BY-SA sources: ARC, SQuAD, DBpedia abstracts, Mr. TyDi Wikipedia passages, BIRD, and 72 tasksource sources flagged share-alike in DATA_SOURCES.md (19,081 rows in v6, 4,328 in v7 warm).

• Copyleft-labelled and possibly non-commercial sources. 12 sources that our later runs drop account for 6,133 rows of v6 (1.3%) and 1,341 rows of v7 warm:

  • the GPL / AGPL / ODbL / MPL-labelled tasksource sources emo/emo2019, dynahate, tomi-nli, multilingual/wili_2018, fever-evidence-related, ethos/multilabel, multilingual/universal-joy, implicatures, hallmarks-of-cancer/yesno and scientific-exaggeration-detection;
  • blog_authorship_corpus (age, gender), which may be non-commercial upstream.

  Both stages were trained before that drop was applied, so both saw these sources.

• Teacher rows. 64,862 rows in v6 and 66,599 in v7 warm. Of v7 warm's, 52,167 are new rows from the R2 snapshot (2026-11-06 12:00 UTC), and the rest are replayed from v6. Logits' pipeline wrote them on its own fleet:

  • Kimi K3.1 Pro designed the task families (Kimi K3.1-Pro License, a modified MIT licence);
  • DeepSeek V4.1 Flash wrote the states (MIT);
  • both models' votes are the labels.

  Data attribution, below, gives both licences' terms.

• DATA_SOURCES.md, a file in this repository, gives each tasksource source, its licence as published and its rows in each stage, and counts the teacher rows. Data attribution, below, credits every other source.

The partners' training data is not fully disclosed:

• KnowLine-31B-Gen1's card describes its data only by example. About a third of its 906,122 rows are Decision Index-format rows (public train splits rewritten, for example WinoGrande XL, ACOS and GSM8K) plus synthetic items. It reports a decontamination against every Decision Index 0.3 row, which we could not verify.
• d3-ultra's authors do not disclose its training data or method.

Credits

• PelaAI, KnowLine-31B-Gen1 @ 8f052a81: Apache-2.0. Member A starts from its weights, which in turn are a fine-tune of google/gemma-4-31B-it.
• vLLM Semantic Router Team, d3-ultra (Decision 3.0) @ 37fee019: Apache-2.0. Member B is its weights, converted to FP8, plus its code, unchanged. Its NOTICE is kept in d3-ultra/NOTICE.
• Google, Gemma 4: Apache-2.0. Gemma is the base of member A. "Gemma" here only describes that origin; Gemma is a trademark of Google LLC, and no endorsement is implied.
• Logits: our LoRA, prompt layout, engine and pair code, under Apache-2.0. code/logits_sigma_1_duo/fp8_kernel.py vendors Apache-2.0 Triton kernels from transformers v5.0.0 (attributed in the file).

Modifications, as Apache-2.0 §4 asks:

• member A's weights are KnowLine's with our LoRA merged in, then converted to FP8;
• member B's weights are d3-ultra's, converted to FP8.

NOTICE lists every upstream notice. UPSTREAMS.json pins every upstream revision. SHA256SUMS covers every file.

Data attribution

Thanks to everyone who builds and maintains these datasets and records. Every source below was reformatted into decision questions and many were modified (fields selected, text clipped, identifiers masked), so none of it should be read as an official copy of its source. The full per-source list of the 411 tasksource sources, with licence and row counts, is in DATA_SOURCES.md.

Collections

• tasksource-jev-typed-decisions, tasksource/tasksource-jev-typed-decisions, built with tasksource (D. Sileo, "tasksource: A Large Collection of NLP tasks with a Structured Dataset Preprocessing Framework", LREC-COLING 2024). Licences per source (CC BY, CC BY-SA, MIT, Apache-2.0, BSD, CC0, GPL, …); only sources tagged for commercial use were read.

Research datasets

Dataset	Reference	Licence
BANKING77	Casanueva et al., 2020	CC BY 4.0
CLINC150 (+ out-of-scope)	Larson et al., 2019	CC BY 3.0
WinoGrande	Sakaguchi et al., 2020	CC BY
ARC (Easy, Challenge)	Clark et al., 2018	CC BY-SA 4.0
CommonsenseQA	Talmor et al., 2019	MIT
GSM8K	Cobbe et al., 2021	MIT
HellaSwag (ActivityNet contexts)	Zellers et al., 2019	MIT
Amazon ESCI	Reddy et al., 2022; Copyright Amazon.com, Inc. or its affiliates	Apache-2.0
ContractNLI	Koreeda and Manning, 2021	CC BY 4.0
iSarcasmEval	Abu Farha et al., 2022	MIT
MedMCQA	Pal et al., 2022	MIT / Apache-2.0
AQuA-RAT	Ling et al., 2017	Apache-2.0
QASC	Khot et al., 2020	CC BY 4.0
CaseHOLD (LexGLUE)	Zheng et al., 2021; Chalkidis et al., 2022	CC BY 4.0 / Apache-2.0
Cosmos QA	Huang et al., 2019	CC BY 4.0
HEAD-QA	Vilares and Gómez-Rodríguez, 2019	MIT
xLAM function calling 60k, Synth-APIGen	Liu et al., 2024 (APIGen); via argilla/apigen-function-calling	CC BY 4.0
ToolACE	Liu et al., 2024	Apache-2.0
glaive-function-calling-v2	Glaive AI	Apache-2.0
WANDS	Chen et al., 2022 (Wayfair)	MIT
DBpedia-Entity v2 (BEIR copy)	Hasibi et al., 2017; Thakur et al., 2021; DBpedia abstracts	MIT (qrels); CC BY-SA 3.0 (abstracts)
Mr. TyDi (English)	Zhang et al., 2021; Wikipedia passages by Wikipedia contributors	Apache-2.0; CC BY-SA 3.0 (text)
SQuAD v1.1	Rajpurkar et al., 2016	CC BY-SA 4.0
BIRD (train)	J. Li et al., "Can LLM Already Serve as A Database Interface? A BIg Bench for Large-Scale Database Grounded Text-to-SQLs", NeurIPS 2023	CC BY-SA 4.0
The Public Jira Dataset	Montgomery L, Lüders C, Maalej W., "An Alternative Issue Tracking Dataset of Public Jira Repositories", MSR 2022; data https://zenodo.org/records/15719919, (c) the authors and the contributors to the 16 public Jira trackers	CC BY 4.0
args.me corpus 1.0-cleaned (debate.org part)	Ajjour, Wachsmuth, Kiesel, Potthast, Hagen, Stein, 2019; https://zenodo.org/records/4139439	CC BY 4.0

Changes, for every CC BY and CC BY-SA source: items were sampled, filtered and rewritten as decision questions with options; long text was clipped; for Jira, e-mail addresses, phone numbers and user mentions were masked and priority, resolution and issue-type names were mapped onto common scales; for BIRD, schemas were trimmed, example rows added, gold queries mutated and executed; for args.me, posts were paired with their own debate's resolution, a target phrase or an unrelated topic. No endorsement by any author is implied. Licences: CC BY 4.0, CC BY 3.0, CC BY-SA 4.0, CC BY-SA 3.0.

Public records

• NHTSA Office of Defects Investigation vehicle complaints, National Highway Traffic Safety Administration (public domain, 17 USC 105).
• OSHA Severe Injury Reports, U.S. Department of Labor, Occupational Safety and Health Administration (public domain).
• NTSB aviation accident database, courtesy of the National Transportation Safety Board (public domain; evaluation only).
• Federal Judicial Center Integrated Database (civil and criminal), https://www.fjc.gov/research/idb (public domain).
• openFDA enforcement reports and device adverse events, https://open.fda.gov/ (CC0 1.0). GMDN terms were not used.
• USPTO. Source: United States Patent and Trademark Office, www.uspto.gov. Trademark data from the USPTO Trademark Case Files Dataset (2023 update) and the USPTO trademark daily application files. Graham, Stuart J.H., Marco, Alan C., Miller, Richard (2018). The USPTO Trademark Case Files Dataset: Descriptions, Lessons, and Insights. Journal of Economics & Management Strategy. https://doi.org/10.1111/jems.12035
• CVE. Copyright (c) 1999-2026, The MITRE Corporation. CVE and the CVE logo are registered trademarks of The MITRE Corporation. Terms of Use (https://www.cve.org/Legal/TermsOfUse): MITRE hereby grants you a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable copyright license to reproduce, prepare derivative works of, publicly display, publicly perform, sublicense, and distribute Common Vulnerabilities and Exposures (CVE). Any copy you make for such purposes is authorized provided that you reproduce MITRE's copyright designation and this license in any such copy.
• CWE. Copyright (c) 2006-2026, The MITRE Corporation. CWE is a trademark of The MITRE Corporation. Terms of Use (https://cwe.mitre.org/about/termsofuse.html): The MITRE Corporation hereby grants you a non-exclusive, royalty-free license to use CWE for research, development, and commercial purposes. Any copy you make for such purposes is authorized on the condition that you reproduce MITRE's copyright designation and this license in any such copy.
• CISA Known Exploited Vulnerabilities catalog, https://github.com/cisagov/kev-data (CC0 1.0). Use does not imply endorsement by CISA or DHS.

The CFPB Consumer Complaint Database narratives and the nebius SWE-agent trajectories are held out for evaluation only; no row of either was trained on.

Teacher-generated data

• Task families: written by Kimi K3.1 Pro (moonshotai/Kimi-K3.1-Pro), served on the Logits fleet. Kimi K3.1-Pro License (a modified MIT licence, https://huggingface.co/moonshotai/Kimi-K3.1-Pro/blob/main/LICENSE): no clause restricts the use of outputs, distillation or training other models. Its attribution clause covers commercial products above 100 million monthly active users or 20 million US dollars of monthly revenue that use the Software or a derivative work; the licence does not settle whether a model trained on outputs is a derivative work.
• States: written by DeepSeek V4.1 Flash (deepseek-ai/DeepSeek-V4.1-Flash), MIT License, Copyright (c) 2023 DeepSeek.
• Labels: vote histograms of both models.

Prompt layout and recipe

Listing a request's questions before the state follows published decision-model work (FlexDM, github.com/1000luca/FlexDM, MIT; Sieve; Turion Jev; Aleria Judge). Reading short prompts twice follows KnowLine Gen1.5 and RE2 (Xu et al., 2023, arXiv:2309.06275). The loss recipe (learning rate 1e-4, KL to the base 1.0, Brier 0.3) follows the FlexDM-LoRA 2.1 model card (1000luca/FlexDM-LoRA-v2.1). No code was copied from them.

Limitations

• Hardware. It needs one ~96 GB GPU, and the FP8 path needs CUDA 13 drivers (≥ 580). It was measured on an H200 and on an RTX PRO 6000 Blackwell.
• Latency is close to the limit. The mean latency is within ~10% of the board's 1,000 ms limit, and requests with many long questions take seconds.
• Very long requests. Above 50,000 characters, only member A answers.
• Score. The index is a probability average of two models. Most of its gain over member A alone comes from d3-ultra.