judgly
Copyright 2026 Timo Lassmann

This product is licensed under the Apache License, Version 2.0 (see LICENSE).

------------------------------------------------------------------------------
Models
------------------------------------------------------------------------------

judgly is designed to run on Google's Gemma 4 models (the default is Gemma 4
12B, downloaded from https://huggingface.co/ggml-org/gemma-4-12B-it-GGUF).
Gemma is provided by Google under the Apache License, Version 2.0; see
https://ai.google.dev/gemma/docs/gemma_4_license. The smaller pack
uses Qwen3-4B-Instruct-2507 by the Qwen team, Alibaba Cloud, under the Apache
License, Version 2.0 (downloaded from
https://huggingface.co/unsloth/Qwen3-4B-Instruct-2507-GGUF). The model weights
are not included in this package: they are downloaded from their publisher on
first use, and their own licence and notices apply to them. "Gemma" is a
trademark of Google LLC and "Qwen" names Alibaba Cloud's models; both are used
here only to describe the model judgly runs on and do not imply endorsement.
Use of Gemma is also subject to Google's Gemma Prohibited Use Policy
(https://ai.google.dev/gemma/prohibited_use_policy).

------------------------------------------------------------------------------
Third-party software
------------------------------------------------------------------------------

The native library libjudgly and the command-line tools are built from the
following third-party code, each under the licence shown. The full licence
texts are in the LICENSES/ directory of the source tree, the sdist and the
wheel.

  llama.cpp and ggml   MIT License   Copyright (c) 2023-2026 The ggml authors
                                     https://github.com/ggml-org/llama.cpp
                                     LICENSES/llama.cpp-ggml-MIT.txt
  llamafile sgemm      MIT License   Copyright 2024 Mozilla Foundation
    (inside ggml)                    LICENSES/llamafile-sgemm-MIT.txt
  YaRN RoPE            MIT License   Copyright (c) 2023 Jeffrey Quesnelle and
    (inside ggml)                    Bowen Peng; LICENSES/yarn-rope-MIT.txt
  yyjson               MIT License   Copyright (c) 2020 YaoYuan
                                     https://github.com/ibireme/yyjson
                                     LICENSES/yyjson-MIT.txt
  ggllm.cpp tokenizer  MIT License   Copyright (c) 2023 https://github.com/cmp-nct
    (inside llama.cpp)               LICENSES/ggllm-cpp-MIT.txt

llama.cpp's BPE tokenizer (src/llama-vocab.cpp) is marked as adapted from
ggllm.cpp (https://github.com/cmp-nct/ggllm.cpp); its licence is in
LICENSES/ggllm-cpp-MIT.txt. llama.cpp is pinned at
commit 6f41ac59e0a49a00483a316a22ada6b04edd2950. Other code vendored in the
llama.cpp tree (for example nlohmann/json, cpp-httplib, stb, miniaudio and the
hash functions under vendor/) is not compiled into libjudgly or the tools.

The C engine in csrc/ and tools/ is written by the author of judgly and is
under this package's licence.

------------------------------------------------------------------------------
Data
------------------------------------------------------------------------------

This version ships calibration heads inside the packs
(src/judgly/packs/<name>/heads/), two per format (the H2 head and the per-type
temperature, fitted on the same data), licensed as stated in the LICENSE file
next to them:

  general heads ("*")  Apache License 2.0. Fitted on MMLU, CommonsenseQA,
                       StrategyQA, HH-RLHF (MIT); QASC, Cosmos QA, BANKING77,
                       HelpSteer2, LEDGAR as released in LexGLUE (CC BY 4.0);
                       Civil Comments (CC0 1.0); GoEmotions (Apache 2.0); and
                       judgly's own generated tasks.
  stance heads         CC BY-SA 4.0 (LICENSES/CC-BY-SA-4.0.txt). Fitted on
                       MultiNLI (OANC licence, parts CC BY-SA 3.0 / CC BY 3.0);
                       VitaminC and FEVER (CC BY-SA 3.0), which contain
                       material derived from Wikipedia (Wikipedia contributors,
                       CC BY-SA 3.0 / 4.0); SNLI (CC BY-SA 4.0), which includes
                       Flickr30k captions; WANLI (CC BY 4.0); and SciNLI
                       (CC BY-SA 4.0).

Heads contain no dataset text. Datasets used only for evaluation are not
redistributed; they are downloaded by scripts/prep_tiers.py from their
publishers. Every dataset, its pinned revision, licence and citation is listed
in data/registry.yaml, and the licence evidence is in docs/licences.md.

------------------------------------------------------------------------------
External comparison
------------------------------------------------------------------------------

docs/results/external-comparison/ holds the answers (item ids, probabilities,
token counts and timings) of three decision models run through Ollama 0.35.0
(https://github.com/ollama/ollama, MIT License):

  Nimble 9B            Bespoke Labs and Maheswaran Sathiamoorthy,
                       https://huggingface.co/bespokelabs/Bespoke-Nimble-9B
                       (weights: Apache License 2.0, as stated on the model card)
  Tev1 4B, Tev1 0.8B   Together AI, https://github.com/togethercomputer/tev1
                       (code: MIT License, Copyright (c) 2026 open-jev
                       contributors; the model cards describe the weights'
                       licence as being finalized)

No weights of these models are included in judgly or in this repository; judgly
does not use them.
