Repository navigation
Expand file tree
/
Copy pathNOTICE
More file actions
47 lines (40 loc) · 1.99 KB
/
Copy pathNOTICE
File metadata and controls
47 lines (40 loc) · 1.99 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
This repository is licensed under the MIT License (see LICENSE), except for
the third-party material listed below.
Attention Matching
------------------
This code base is derived from https://github.com/adamzweiger/compaction
(Zweiger et al., "Fast KV Compaction via Attention Matching", arXiv:2602.16284),
MIT License, Copyright (c) 2026 Adam Zweiger.
Hugging Face Transformers
-------------------------
The following files are modified copies of files from Hugging Face
Transformers 4.57.1 (https://github.com/huggingface/transformers), licensed
under the Apache License 2.0 (see LICENSES/Apache-2.0.txt). Each file keeps its
original copyright header and states how it was modified.
models/llama/modeling_llama.py
models/qwen3/modeling_qwen3.py
models/qwen3_moe/modeling_qwen3_moe.py
models/gemma3/modeling_gemma3.py
Cartridges
----------
compaction/query_generation/self_study.py is adapted from Cartridges
(https://github.com/HazyResearch/cartridges), licensed under the Apache
License 2.0 (see LICENSES/Apache-2.0.txt).
Data
----
data/QuALITY.v1.0.1.htmlstripped.dev is part of QuALITY
(Pang et al., 2022, https://github.com/nyu-mll/quality), distributed under the
Creative Commons Attribution 4.0 International License (CC BY 4.0,
https://creativecommons.org/licenses/by/4.0/). The file is unmodified.
LongHealth and Long Code Arena are not included. They are downloaded
(LongHealth) or built (scripts/build_lca_repo_completion.py) from their
original sources, under their own licenses.
Model-derived files
-------------------
The per-head budget files (head_budget_optimization/head_budgets/) and the
Q-Filters vectors (qfilters/pile/) were computed from the activations of
Qwen3-4B, Qwen3-4B-Instruct-2507, Qwen3-30B-A3B (Apache License 2.0),
Llama-3.1-8B-Instruct and Gemma-3-12B-it. Built with Llama.
Llama 3.1 is licensed under the Llama 3.1 Community License, Copyright (c)
Meta Platforms, Inc. All Rights Reserved. Gemma is provided under and subject
to the Gemma Terms of Use found at ai.google.dev/gemma/terms.