indahl commited on
Commit
bf8ced5
·
verified ·
1 Parent(s): f9e38e2

Data pack as committed at indahl.ai hfsize/hfsize-data-pack.zip

Browse files
README.md ADDED
@@ -0,0 +1,64 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ pretty_name: "HFSIZE Data Pack"
3
+ license: other
4
+ license_name: per-part-see-license
5
+ license_link: https://huggingface.co/datasets/indahl/hfsize-data-pack/blob/main/pack/LICENSE.txt
6
+ tags:
7
+ - indahl
8
+ - heuremen
9
+ - hfsize
10
+ - stated-vs-delivered
11
+ configs:
12
+ - config_name: hfsize_outcomes
13
+ default: true
14
+ data_files: pack/results/hfsize_outcomes.csv
15
+ - config_name: hfsize_summary
16
+ data_files: pack/results/hfsize_summary.csv
17
+ ---
18
+ # HFSIZE Data Pack
19
+
20
+ Heuremen LLC · https://indahl.ai/hfsize
21
+
22
+ As of 15 Nov 2026 04:12, the time stamped on the desk's feed when this pack was built.
23
+
24
+ ## The question
25
+
26
+ Model names against the weights. The page states it in full at https://indahl.ai/hfsize.
27
+
28
+ ## The ruler
29
+
30
+ Agree: the weights hold between 0.70 and 1.35 times the stated size. Disagree: outside that band. Above: the lower end of the 95 percent interval of the disagreement share is above 0.10. Consistent: the upper end is below 0.05. Otherwise, and with fewer than 30 comparable models, cannot tell. The list of models where Hugging Face's own total contradicts its own breakdown was defined after the first read and carries no outcome. Source: the Hugging Face API.
31
+
32
+ ## Four outcomes
33
+
34
+ ABOVE, BELOW, CONSISTENT and CANNOT_TELL are the four outcomes. A row reads CANNOT_TELL when its interval does not settle the question. Where the page uses another word, such as NOTHING_RETURNED, it is kept as printed. The lo and hi columns are the interval bounds the desk prints with each result, and the page names the method.
35
+
36
+ results/hfsize_outcomes.csv holds 1 row: 1 ABOVE.
37
+
38
+ ## Files
39
+
40
+ - results/hfsize_summary.json: our counts, outcomes, intervals and labels from the feed, with every other value left out.
41
+ - results/hfsize_summary.csv: the same file flattened to one row per value, with its path.
42
+ - results/hfsize_outcomes.json and results/hfsize_outcomes.csv: every result in it that carries an outcome or an interval, one row each, with path, label, outcome, n, of, share, lo and hi columns.
43
+ - SOURCES.md: every source with its address, licence and pull date.
44
+ - LICENSE.txt and NOTICE.txt: the terms of each part and the credit lines.
45
+ - CITATION.cff: how to cite this pack.
46
+ - TESTED.md: what was checked against the live page and feed before this pack was placed.
47
+ - SHA256SUMS.txt: the sha256 of every other file in the pack.
48
+
49
+ ## Licence
50
+
51
+ This pack carries only our derived results (counts, outcomes and intervals) and links to the sources, because the Hugging Face API is under its terms of service and not an open licence.
52
+
53
+ ## Cite
54
+
55
+ Heuremen LLC, IRIS Desk Packs, Zenodo, doi:10.5281/zenodo.24110664
56
+
57
+ https://doi.org/10.5281/zenodo.24110664
58
+
59
+ This pack is one file of that Zenodo record, which holds the packs of the desks listed there. That DOI always points to the newest version on Zenodo. Each version there also has its own DOI.
60
+
61
+
62
+ ---
63
+
64
+ The pack is in pack/ exactly as published at https://indahl.ai/hfsize/hfsize-data-pack.zip; pack/SHA256SUMS.txt verifies every file. Paths above are relative to pack/. Terms: pack/LICENSE.txt and pack/NOTICE.txt.
pack/CITATION.cff ADDED
@@ -0,0 +1,10 @@
 
 
 
 
 
 
 
 
 
 
 
1
+ cff-version: 1.2.0
2
+ message: "If you use this data pack, please cite it. Terms are given in LICENSE.txt."
3
+ title: "HFSIZE Data Pack: Model names against the weights"
4
+ authors:
5
+ - name: "Heuremen LLC"
6
+ version: "2026-11-15"
7
+ date-released: "2026-11-15"
8
+ doi: "10.5281/zenodo.24110664"
9
+ url: "https://indahl.ai/hfsize"
10
+ repository-code: "https://github.com/Heuremen-LLC/indahl.ai"
pack/LICENSE.txt ADDED
@@ -0,0 +1,13 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ HFSIZE Data Pack
2
+ Heuremen LLC · https://indahl.ai/hfsize
3
+
4
+ This pack mixes material from several sources. Each part keeps the terms of its source, and Heuremen LLC adds no restriction to any of them. The parts are listed here.
5
+
6
+ 1. Heuremen LLC's own parts
7
+
8
+ The README, SOURCES.md, NOTICE.txt, LICENSE.txt, TESTED.md, CITATION.cff, SHA256SUMS.txt, the build code, the compilation (the flattening, the CSV conversion, the column names), and the counts, outcomes, intervals and labels written by Heuremen LLC. Every file under results/ in this pack is one of these parts.
9
+
10
+ Heuremen LLC permits anyone to use, copy and change these parts with attribution to Heuremen LLC and https://indahl.ai/hfsize. For its own text (the README, SOURCES.md, NOTICE.txt, LICENSE.txt, TESTED.md and the labels written by Heuremen LLC) that permission does not include use as AI training data.
11
+
12
+ Heuremen LLC reserves text and data mining and use as AI training data for that text only. That reservation does not apply to values that come from the sources.
13
+
pack/NOTICE.txt ADDED
@@ -0,0 +1,13 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ HFSIZE Data Pack
2
+ Heuremen LLC · https://indahl.ai/hfsize
3
+
4
+ Attribution. The full terms are in LICENSE.txt and this notice says the same.
5
+
6
+ Sources (linked, not copied)
7
+ Hugging Face Hub API. https://huggingface.co/docs/hub/api . Hugging Face terms of service, https://huggingface.co/terms-of-service.
8
+
9
+ Heuremen LLC
10
+ The README, SOURCES.md, NOTICE.txt, LICENSE.txt, TESTED.md, CITATION.cff, SHA256SUMS.txt, the build code, the compilation (the flattening, the CSV conversion, the column names), and the counts, outcomes, intervals and labels written by Heuremen LLC may be used, copied and changed with attribution to Heuremen LLC and https://indahl.ai/hfsize.
11
+ For the text files and the labels written by Heuremen LLC, that permission does not include use as AI training data. The reservation covers that text only.
12
+
13
+ To correct or remove a row, use https://indahl.ai/correction.
pack/README.md ADDED
@@ -0,0 +1,42 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # HFSIZE Data Pack
2
+
3
+ Heuremen LLC · https://indahl.ai/hfsize
4
+
5
+ As of 15 Nov 2026 04:12, the time stamped on the desk's feed when this pack was built.
6
+
7
+ ## The question
8
+
9
+ Model names against the weights. The page states it in full at https://indahl.ai/hfsize.
10
+
11
+ ## The ruler
12
+
13
+ Agree: the weights hold between 0.70 and 1.35 times the stated size. Disagree: outside that band. Above: the lower end of the 95 percent interval of the disagreement share is above 0.10. Consistent: the upper end is below 0.05. Otherwise, and with fewer than 30 comparable models, cannot tell. The list of models where Hugging Face's own total contradicts its own breakdown was defined after the first read and carries no outcome. Source: the Hugging Face API.
14
+
15
+ ## Four outcomes
16
+
17
+ ABOVE, BELOW, CONSISTENT and CANNOT_TELL are the four outcomes. A row reads CANNOT_TELL when its interval does not settle the question. Where the page uses another word, such as NOTHING_RETURNED, it is kept as printed. The lo and hi columns are the interval bounds the desk prints with each result, and the page names the method.
18
+
19
+ results/hfsize_outcomes.csv holds 1 row: 1 ABOVE.
20
+
21
+ ## Files
22
+
23
+ - results/hfsize_summary.json: our counts, outcomes, intervals and labels from the feed, with every other value left out.
24
+ - results/hfsize_summary.csv: the same file flattened to one row per value, with its path.
25
+ - results/hfsize_outcomes.json and results/hfsize_outcomes.csv: every result in it that carries an outcome or an interval, one row each, with path, label, outcome, n, of, share, lo and hi columns.
26
+ - SOURCES.md: every source with its address, licence and pull date.
27
+ - LICENSE.txt and NOTICE.txt: the terms of each part and the credit lines.
28
+ - CITATION.cff: how to cite this pack.
29
+ - TESTED.md: what was checked against the live page and feed before this pack was placed.
30
+ - SHA256SUMS.txt: the sha256 of every other file in the pack.
31
+
32
+ ## Licence
33
+
34
+ This pack carries only our derived results (counts, outcomes and intervals) and links to the sources, because the Hugging Face API is under its terms of service and not an open licence.
35
+
36
+ ## Cite
37
+
38
+ Heuremen LLC, IRIS Desk Packs, Zenodo, doi:10.5281/zenodo.24110664
39
+
40
+ https://doi.org/10.5281/zenodo.24110664
41
+
42
+ This pack is one file of that Zenodo record, which holds the packs of the desks listed there. That DOI always points to the newest version on Zenodo. Each version there also has its own DOI.
pack/SHA256SUMS.txt ADDED
@@ -0,0 +1,10 @@
 
 
 
 
 
 
 
 
 
 
 
1
+ 73e3f84df1cc76ff5ec3698de311be777ab13365771b898d866cedcbff115a6e CITATION.cff
2
+ c402adb8854d0173aed46bbf6bd97de218638522846276eb8b257202390888a0 LICENSE.txt
3
+ 884efba2ee1e78954c3be53527a5e7fdc9b79a217cd1e26d47519eeb79e57d27 NOTICE.txt
4
+ 9fd8cee564bcaafd29a75d23c7ef60303849dc73f2608c6c424a9870dcd917c6 README.md
5
+ 71cdf7b4f986f63cfcd54b0cd32fbfe1f4f88968ac859f6c2fc7318eb9f07fe2 SOURCES.md
6
+ 9cdfc1ca9ebc7c71cd7afd0c5f9a548c34653aca8d0fd181390f8475221c65fd TESTED.md
7
+ fd7f16e9ff17e10d6859e1807a07ad84191f4e64c1dcc51ead68f14c139da4e1 results/hfsize_outcomes.csv
8
+ e1dc7c14f52f4b538c007ce289823b26c9331ed2a8d177c6d475e5dd54b1d1e7 results/hfsize_outcomes.json
9
+ f772bb371d4493296f593dc10b728381e01896bb3f35763242b9098a30e19aca results/hfsize_summary.csv
10
+ 4af5e45a6e2f55edb89b4c541d548bfa82b7b39a26dff796fc2a3809154d5466 results/hfsize_summary.json
pack/SOURCES.md ADDED
@@ -0,0 +1,7 @@
 
 
 
 
 
 
 
 
1
+ # HFSIZE Data Pack: sources
2
+
3
+ Each source is read by the desk on its own schedule. The feed in this pack was read at 15 Nov 2026 04:12, and that is the pull date for every source below.
4
+
5
+ | Source | Address | Licence or terms | Pulled |
6
+ |---|---|---|---|
7
+ | Hugging Face Hub API | https://huggingface.co/docs/hub/api | Hugging Face terms of service, https://huggingface.co/terms-of-service | 15 Nov 2026 04:12 |
pack/TESTED.md ADDED
@@ -0,0 +1,10 @@
 
 
 
 
 
 
 
 
 
 
 
1
+ # HFSIZE Data Pack: tested
2
+
3
+ Every CSV in this pack was opened again with a CSV reader on 2026-11-15 05:02 UTC and checked against the live feed at https://indahl.ai/hfsize/live.json.
4
+
5
+ - holds: hfsize_summary.csv: 71 rows, one per value of the live feed (71)
6
+ - holds: hfsize_outcomes.csv: 1 rows, the live feed holds 1 results, outcome, n, lo and hi equal on every row
7
+ - holds: page headline "12 of 134": 12 is in the pack
8
+ - holds: page headline "12 of 134": 134 is in the pack
9
+
10
+ TESTED 2026-11-15 05:02 UTC: 4 checks, every one holds.
pack/results/hfsize_outcomes.csv ADDED
@@ -0,0 +1,2 @@
 
 
 
1
+ path,label,outcome,n,of,share,lo,hi,share_lo,share_hi,reason,organizations,awaiting
2
+ /comparison,comparison,ABOVE,12,134,0.0896,0.052,0.15,0.052,0.15,,7,0
pack/results/hfsize_outcomes.json ADDED
@@ -0,0 +1,17 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "path": "/comparison",
4
+ "label": "comparison",
5
+ "outcome": "ABOVE",
6
+ "n": 12,
7
+ "of": 134,
8
+ "share": 0.0896,
9
+ "lo": 0.052,
10
+ "hi": 0.15,
11
+ "share_lo": 0.052,
12
+ "share_hi": 0.15,
13
+ "reason": null,
14
+ "organizations": 7,
15
+ "awaiting": 0
16
+ }
17
+ ]
pack/results/hfsize_summary.csv ADDED
@@ -0,0 +1,72 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ path,value
2
+ /updated,15 Nov 04:12
3
+ /gate,passed
4
+ /models,1000
5
+ /comparison/of,134
6
+ /comparison/n,12
7
+ /comparison/share,0.0896
8
+ /comparison/share_lo,0.052
9
+ /comparison/share_hi,0.15
10
+ /comparison/outcome,ABOVE
11
+ /comparison/reason,
12
+ /comparison/req13/nothing_returned,0
13
+ /comparison/req13/n_readings,1
14
+ /comparison/req13/cannot_tell,0
15
+ /comparison/organizations,7
16
+ /comparison/awaiting,0
17
+ /classes/AGREE/n,122
18
+ /classes/DISAGREE/n,12
19
+ /classes/NOT_COMPARABLE/n,866
20
+ /not_comparable_reasons/no size in the name/n,476
21
+ /not_comparable_reasons/packed or quantized/n,350
22
+ /not_comparable_reasons/mixture-of-experts name/n,21
23
+ /not_comparable_reasons/no safetensors total/n,15
24
+ /not_comparable_reasons/Hugging Face total does not match its own breakdown/n,4
25
+ /api_mismatch[0]/repo,ornith-ai/Ornith-1.0-35B
26
+ /api_mismatch[0]/total,664944
27
+ /api_mismatch[0]/breakdown,35107181936
28
+ /api_mismatch[1]/repo,farbodtavakkoli/OTel-2.0-LLM-31B-IT-QLoRA
29
+ /api_mismatch[1]/total,31273086512
30
+ /api_mismatch[1]/breakdown,31273088876
31
+ /api_mismatch[2]/repo,ornith-ai/Ornith-1.0-9B
32
+ /api_mismatch[2]/total,1469680
33
+ /api_mismatch[2]/breakdown,9409813744
34
+ /api_mismatch[3]/repo,Infomaniak-AI/vllm-translategemma-4b-it
35
+ /api_mismatch[3]/total,4971331952
36
+ /api_mismatch[3]/breakdown,4300079472
37
+ /disagree_list[0]/repo,Qwen/Qwen3-ASR-1.7B
38
+ /disagree_list[0]/ratio,1.382
39
+ /disagree_list[0]/reason,One set of weights holds an audio encoder and a text decoder (configuration: audio_config and text_config). The card does not give the split.
40
+ /disagree_list[1]/repo,meta-llama/Prompt-Guard-86M
41
+ /disagree_list[1]/ratio,3.242
42
+ /disagree_list[1]/reason,"The name states 86M and the weights hold 278.8M. The configuration is gated, so the split could not be read and the reason is not established."
43
+ /disagree_list[2]/repo,Qwen/Qwen3-TTS-12Hz-0.6B-Base
44
+ /disagree_list[2]/ratio,1.524
45
+ /disagree_list[2]/reason,The configuration holds a talker and a code predictor. The repository carries a second weight file for the speech tokenizer.
46
+ /disagree_list[3]/repo,Qwen/Qwen3-TTS-12Hz-0.6B-CustomVoice
47
+ /disagree_list[3]/ratio,1.51
48
+ /disagree_list[3]/reason,The configuration holds a talker and a code predictor. The repository carries a second weight file for the speech tokenizer.
49
+ /disagree_list[4]/repo,EleutherAI/pythia-70m-deduped
50
+ /disagree_list[4]/ratio,1.366
51
+ /disagree_list[4]/reason,The card's table gives the size in non-embedding parameters. The weights also hold the embeddings.
52
+ /disagree_list[5]/repo,openbmb/MiniCPM-V-4.7-1B
53
+ /disagree_list[5]/ratio,1.554
54
+ /disagree_list[5]/reason,"One set of weights holds a SigLIP2 vision encoder (27 layers) and a Qwen3.5 language model (24 layers, embeddings not tied, vocabulary 248,094) with one multi-token-prediction layer (configuration: vision_config and text_config). The card does not give the split."
55
+ /disagree_list[6]/repo,Qwen/Qwen3-ASR-0.6B
56
+ /disagree_list[6]/ratio,1.563
57
+ /disagree_list[6]/reason,One set of weights holds an audio encoder and a text decoder (configuration: audio_config and text_config). The card does not give the split.
58
+ /disagree_list[7]/repo,Qwen/Qwen3-ForcedAligner-0.6B
59
+ /disagree_list[7]/ratio,1.53
60
+ /disagree_list[7]/reason,One set of weights holds an audio encoder and a text decoder (configuration: audio_config and text_config). The card does not give the split.
61
+ /disagree_list[8]/repo,openbmb/MiniCPM5-2B-DSpark
62
+ /disagree_list[8]/ratio,0.162
63
+ /disagree_list[8]/reason,The card describes a draft checkpoint made to pair with MiniCPM5-2B. The name states the size of the model it pairs with.
64
+ /disagree_list[9]/repo,ornith-ai/Ornith-1.5-9B-DFlash
65
+ /disagree_list[9]/ratio,0.144
66
+ /disagree_list[9]/reason,The card describes a draft checkpoint for speculative decoding with Ornith-1.5-9B. The name states the size of the model it pairs with.
67
+ /disagree_list[10]/repo,Salesforce/blip2-opt-2.7b
68
+ /disagree_list[10]/ratio,1.387
69
+ /disagree_list[10]/reason,"The card says the model joins an image encoder, a querying transformer and the 2.7B language model. The name states the language model's size."
70
+ /disagree_list[11]/repo,incoai/Qwen3.8-27B-DFlash2
71
+ /disagree_list[11]/ratio,0.071
72
+ /disagree_list[11]/reason,The card says it is not a standalone language model but a draft model for Qwen3.8-27B. The name states the size of the model it pairs with.
pack/results/hfsize_summary.json ADDED
@@ -0,0 +1,133 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "updated": "15 Nov 04:12",
3
+ "gate": "passed",
4
+ "models": 1000,
5
+ "comparison": {
6
+ "of": 134,
7
+ "n": 12,
8
+ "share": 0.0896,
9
+ "share_lo": 0.052,
10
+ "share_hi": 0.15,
11
+ "outcome": "ABOVE",
12
+ "reason": null,
13
+ "req13": {
14
+ "nothing_returned": 0,
15
+ "n_readings": 1,
16
+ "cannot_tell": 0
17
+ },
18
+ "organizations": 7,
19
+ "awaiting": 0
20
+ },
21
+ "classes": {
22
+ "AGREE": {
23
+ "n": 122
24
+ },
25
+ "DISAGREE": {
26
+ "n": 12
27
+ },
28
+ "NOT_COMPARABLE": {
29
+ "n": 866
30
+ }
31
+ },
32
+ "not_comparable_reasons": {
33
+ "no size in the name": {
34
+ "n": 476
35
+ },
36
+ "packed or quantized": {
37
+ "n": 350
38
+ },
39
+ "mixture-of-experts name": {
40
+ "n": 21
41
+ },
42
+ "no safetensors total": {
43
+ "n": 15
44
+ },
45
+ "Hugging Face total does not match its own breakdown": {
46
+ "n": 4
47
+ }
48
+ },
49
+ "api_mismatch": [
50
+ {
51
+ "repo": "ornith-ai/Ornith-1.0-35B",
52
+ "total": 664944,
53
+ "breakdown": 35107181936
54
+ },
55
+ {
56
+ "repo": "farbodtavakkoli/OTel-2.0-LLM-31B-IT-QLoRA",
57
+ "total": 31273086512,
58
+ "breakdown": 31273088876
59
+ },
60
+ {
61
+ "repo": "ornith-ai/Ornith-1.0-9B",
62
+ "total": 1469680,
63
+ "breakdown": 9409813744
64
+ },
65
+ {
66
+ "repo": "Infomaniak-AI/vllm-translategemma-4b-it",
67
+ "total": 4971331952,
68
+ "breakdown": 4300079472
69
+ }
70
+ ],
71
+ "disagree_list": [
72
+ {
73
+ "repo": "Qwen/Qwen3-ASR-1.7B",
74
+ "ratio": 1.382,
75
+ "reason": "One set of weights holds an audio encoder and a text decoder (configuration: audio_config and text_config). The card does not give the split."
76
+ },
77
+ {
78
+ "repo": "meta-llama/Prompt-Guard-86M",
79
+ "ratio": 3.242,
80
+ "reason": "The name states 86M and the weights hold 278.8M. The configuration is gated, so the split could not be read and the reason is not established."
81
+ },
82
+ {
83
+ "repo": "Qwen/Qwen3-TTS-12Hz-0.6B-Base",
84
+ "ratio": 1.524,
85
+ "reason": "The configuration holds a talker and a code predictor. The repository carries a second weight file for the speech tokenizer."
86
+ },
87
+ {
88
+ "repo": "Qwen/Qwen3-TTS-12Hz-0.6B-CustomVoice",
89
+ "ratio": 1.51,
90
+ "reason": "The configuration holds a talker and a code predictor. The repository carries a second weight file for the speech tokenizer."
91
+ },
92
+ {
93
+ "repo": "EleutherAI/pythia-70m-deduped",
94
+ "ratio": 1.366,
95
+ "reason": "The card's table gives the size in non-embedding parameters. The weights also hold the embeddings."
96
+ },
97
+ {
98
+ "repo": "openbmb/MiniCPM-V-4.7-1B",
99
+ "ratio": 1.554,
100
+ "reason": "One set of weights holds a SigLIP2 vision encoder (27 layers) and a Qwen3.5 language model (24 layers, embeddings not tied, vocabulary 248,094) with one multi-token-prediction layer (configuration: vision_config and text_config). The card does not give the split."
101
+ },
102
+ {
103
+ "repo": "Qwen/Qwen3-ASR-0.6B",
104
+ "ratio": 1.563,
105
+ "reason": "One set of weights holds an audio encoder and a text decoder (configuration: audio_config and text_config). The card does not give the split."
106
+ },
107
+ {
108
+ "repo": "Qwen/Qwen3-ForcedAligner-0.6B",
109
+ "ratio": 1.53,
110
+ "reason": "One set of weights holds an audio encoder and a text decoder (configuration: audio_config and text_config). The card does not give the split."
111
+ },
112
+ {
113
+ "repo": "openbmb/MiniCPM5-2B-DSpark",
114
+ "ratio": 0.162,
115
+ "reason": "The card describes a draft checkpoint made to pair with MiniCPM5-2B. The name states the size of the model it pairs with."
116
+ },
117
+ {
118
+ "repo": "ornith-ai/Ornith-1.5-9B-DFlash",
119
+ "ratio": 0.144,
120
+ "reason": "The card describes a draft checkpoint for speculative decoding with Ornith-1.5-9B. The name states the size of the model it pairs with."
121
+ },
122
+ {
123
+ "repo": "Salesforce/blip2-opt-2.7b",
124
+ "ratio": 1.387,
125
+ "reason": "The card says the model joins an image encoder, a querying transformer and the 2.7B language model. The name states the language model's size."
126
+ },
127
+ {
128
+ "repo": "incoai/Qwen3.8-27B-DFlash2",
129
+ "ratio": 0.071,
130
+ "reason": "The card says it is not a standalone language model but a draft model for Qwen3.8-27B. The name states the size of the model it pairs with."
131
+ }
132
+ ]
133
+ }