thaki-AI commited on
Commit
9656101
·
verified ·
1 Parent(s): 0a47ce9

Add source license audit (repo IDs, revisions, licenses, citations) to README

Browse files
Files changed (2) hide show
  1. MANIFEST.sha256 +1 -1
  2. README.md +58 -6
MANIFEST.sha256 CHANGED
@@ -1,4 +1,4 @@
1
- 37fa3230d0e1b477153df6881a9ee6b946d39d44c9d6f5850a7c2f2e3b46a5d9 README.md
2
  47be8efd7e1fd31c40bba19c677eed4ced535acdd522f428db29e11aacb39f5d REPRODUCE.md
3
  71dc6a4089b81a02634fa40095c900c30173716ec062db4e118fe8a7a66fbec7 data/final81_arms.csv
4
  1cf6dea059f599d82be869d5b18db384fc03397630175f3b9202383bb3b09dce data/final81_axes.csv
 
1
+ 2fab86a9e538abeda0c3b7d9c4a3bef5d906d5fd59ba8155d73faaa8e6c228df README.md
2
  47be8efd7e1fd31c40bba19c677eed4ced535acdd522f428db29e11aacb39f5d REPRODUCE.md
3
  71dc6a4089b81a02634fa40095c900c30173716ec062db4e118fe8a7a66fbec7 data/final81_arms.csv
4
  1cf6dea059f599d82be869d5b18db384fc03397630175f3b9202383bb3b09dce data/final81_axes.csv
README.md CHANGED
@@ -29,10 +29,13 @@ configs:
29
 
30
  Aggregate results, recipes and serving measurements from the paper
31
  *Beyond Bitwidth: Deployment-Constrained Mixed-Precision Quantization for Multilingual LLM Inference on NVIDIA Blackwell*
32
- (arXiv preprint in preparation; the link will be added once it is public).
33
 
34
- This repository holds **results only**. It contains no model weights, no evaluation items, no per-item predictions
35
- and no calibration text. The quantized checkpoints evaluated in the paper were not retained.
 
 
 
36
 
37
  A separate build of the A81 recipe is released as weights in
38
  [ThakiCloud/Qwen3-8B-NVFP4A16-INT8-Mixed-B200](https://huggingface.co/ThakiCloud/Qwen3-8B-NVFP4A16-INT8-Mixed-B200).
@@ -115,9 +118,58 @@ Differences are accuracy points on the same items; intervals are simultaneous at
115
 
116
  ## Data sources and license
117
 
118
- The evaluation sets were assembled from public benchmarks (KMMLU, KoBEST, MMLU, ARC, HellaSwag, BBH, BIG-bench,
119
- MuSR, HRM8K) and a programmatic numeric generator. This repository republishes none of their items.
120
- The aggregate results in this repository are released under CC BY 4.0 by ThakiCloud.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
121
  `Qwen3-8B` is distributed by its authors under Apache-2.0.
122
 
123
  ## Citation
 
29
 
30
  Aggregate results, recipes and serving measurements from the paper
31
  *Beyond Bitwidth: Deployment-Constrained Mixed-Precision Quantization for Multilingual LLM Inference on NVIDIA Blackwell*
32
+ (submitted to arXiv; the link will be added once it is public).
33
 
34
+ This repository holds **results only**: the record of the D81 experiment (aggregate accuracy, paired differences,
35
+ kernel checks, serving measurements and the A81 precision allocation). It is **not** the 5,168-item FINAL-81
36
+ benchmark or any other evaluation set, and it cannot be used to rerun the evaluation on its own. It contains no model
37
+ weights, no evaluation items, no per-item predictions and no calibration text. The quantized checkpoints evaluated in
38
+ the paper were not retained.
39
 
40
  A separate build of the A81 recipe is released as weights in
41
  [ThakiCloud/Qwen3-8B-NVFP4A16-INT8-Mixed-B200](https://huggingface.co/ThakiCloud/Qwen3-8B-NVFP4A16-INT8-Mixed-B200).
 
118
 
119
  ## Data sources and license
120
 
121
+ ### What this repository contains
122
+
123
+ Every file here is a measurement or a setting produced by ThakiCloud: per-arm and per-axis accuracy aggregates,
124
+ paired differences with intervals, kernel assignments, serving latency, gate thresholds, set hashes and the A81
125
+ allocation. These files are released under CC BY 4.0. They contain no question, passage, answer option, gold label,
126
+ generated output, per-item score or calibration text from any source dataset. A re-audit on 2026-11-17 scanned all
127
+ 9 published files (392 string values) and found no item text and no item-keyed field. No AI-Hub data was used.
128
+
129
+ ### Sources of the evaluation sets (not redistributed)
130
+
131
+ FINAL-81 (5,168 items), SEL-81 (5,052) and DIAG-81 (1,998) were built by a script that pins each source to a Hub
132
+ revision. The license column is what the dataset card declares at that revision.
133
+
134
+ | Source | Hugging Face repo | Revision | License (Hub card) | D81 axis | Citation |
135
+ |---|---|---|---|---|---|
136
+ | KMMLU | `HAERAE-HUB/KMMLU` | `d61b3f19e552c576bf5960dd24289763edc36a88` | CC BY-ND 4.0 | Korean | Son et al., NAACL 2025 (arXiv:2402.11548) |
137
+ | KoBEST (BoolQ, COPA, HellaSwag, SentiNeg; test) | `skt/kobest_v1` | `a5ea15e3ac77ed694b79f6204eb31889a2ba989f` | CC BY-SA 4.0 | Korean in-domain (report-only) | Jang et al., COLING 2022 (arXiv:2204.04541) |
138
+ | MMLU (test) | `cais/mmlu` | `c30699e8356da336a370243923dbaf21066bb9fe` | MIT | English | Hendrycks et al., ICLR 2021 (arXiv:2009.03300) |
139
+ | ARC (Challenge, Easy; test) | `allenai/ai2_arc` | `210d026faf9955653af8916fad021475a3f00453` | CC BY-SA 4.0 | English | Clark et al., 2018 (arXiv:1803.05457) |
140
+ | HellaSwag (validation) | `Rowan/hellaswag` | `218ec52e09a7e7462a5400043bb9a69a41d06b76` | none declared (see notes) | English | Zellers et al., ACL 2019 (arXiv:1905.07830) |
141
+ | BIG-Bench Hard | `lukaemon/bbh` | `982bb89fd79532a8ac676a61fc42eb1aeec63f99` | none declared (see notes) | Reasoning | Suzgun et al., Findings of ACL 2023 (arXiv:2210.09261) |
142
+ | BIG-bench (41 tasks) | `tasksource/bigbench` | `b8ee34693560a13a9feac6aa5540449e18b7eddb` | Apache-2.0 | Reasoning | Srivastava et al., TMLR 2023 (arXiv:2206.04615) |
143
+ | MuSR | `TAUR-Lab/MuSR` | `7c365b439a222150f317764d4f16ae6c96d7d94a` | CC BY 4.0 | Reasoning | Sprague et al., ICLR 2024 (arXiv:2310.16049) |
144
+ | HRM8K (`gsm8k_test`, `ksm_test`) | `HAERAE-HUB/HRM8K` | `c360cabf8d733a82455565358b3dc965aab9ba8d` | MIT | Korean math (report-only) | Ko et al., 2025 (arXiv:2501.02448) |
145
+ | Numeric generator | ThakiCloud code, seeds 8101, 8102, 8103 | — | no third-party text | Numeric | — |
146
+
147
+ ### Redistribution conditions
148
+
149
+ This repository publishes none of the material above, so its CC BY 4.0 license covers only our own measurements.
150
+ If you rebuild the sets with `REPRODUCE.md`, download each source yourself and follow its license:
151
+
152
+ - **KMMLU (CC BY-ND 4.0).** You may share the dataset unchanged with attribution, but you may not share adapted
153
+ material. Reformatted prompt files, subsets in our item schema and per-item outputs can count as adapted material.
154
+ Do not redistribute rebuilt KMMLU files.
155
+ - **KoBEST and ARC (CC BY-SA 4.0).** Adapted material must credit the source and be shared under CC BY-SA 4.0. Rebuilt
156
+ KoBEST or ARC items cannot be relicensed under CC BY 4.0 together with the files here.
157
+ - **MuSR (CC BY 4.0), MMLU and HRM8K (MIT), BIG-bench (Apache-2.0).** Attribution and the license notice are required
158
+ when redistributing.
159
+
160
+ This is a summary of the declared licenses, not legal advice.
161
+
162
+ ### Notes from the license check (2026-11-17)
163
+
164
+ - **HellaSwag.** The Hub mirror declares no license at the pinned revision. The original GitHub repository
165
+ (`rowanz/hellaswag`) was blocked when checked, following a DMCA notice filed by wikiHow on 2026-09-14 (recorded in
166
+ `github/dmca`). HellaSwag contexts include text derived from wikiHow articles. We treat redistribution of HellaSwag
167
+ items as unresolved; this repository contains none.
168
+ - **BIG-Bench Hard.** The Hub mirror declares no license, and no license file was detected in the upstream repository
169
+ (`suzgun/BIG-Bench-Hard`). Its tasks come from BIG-bench (Apache-2.0); we did not establish a separate license for BBH.
170
+ - **DEV-80.** Two rows of the paired-comparison table use DEV-80, a set from the earlier D80 stage built by an earlier
171
+ version of the same script. Its source list was not re-audited for this README.
172
+
173
  `Qwen3-8B` is distributed by its authors under Apache-2.0.
174
 
175
  ## Citation