querypieai commited on
Commit
ddefbf0
·
verified ·
1 Parent(s): 6438a35

Release self-contained QEV multitask model

Browse files
MANIFEST.sha256 CHANGED
@@ -10,11 +10,11 @@ f1bc8adc0fca1a1eae1e723116151410b0042b83f97b223aa8bde419037d11c6 FINETUNING.ko.
10
  8d73ef0374c47be1033c1e78cc1427af2cce62c263da3dc06bdb06512d38e2d6 LICENSES/._Apache-2.0.txt
11
  cfc7749b96f63bd31c3c42b5c471bf756814053e847c10f3eb003417bc523d30 LICENSES/Apache-2.0.txt
12
  4b39d0a3e6ff27421e95e7b67378331cc88dd7e2014eb4175f8a5f1c6ee73198 LICENSE_POLICY.md
13
- 106fd4a3220bd6cfdf45ab7f9e8f82e8885d30e586d728b62e2c94cc4d191eba MODEL_LINEAGE.json
14
  baceafe0782b0655c6147f33d369152bbd084d7d3c372afd4c484384485c029d NOTICE
15
- c60ace0fdb167b85f813dd8aef635096a1885c8b7af2a87214c0a89497e245f7 QUANTIZATION.ja.md
16
- 245d4e4039fc5a442f33645130a391f3776ed9317dadd46b933a577c27f241fd QUANTIZATION.ko.md
17
- 7c35622022c260c9e5c1c052fb33ac34dad29b0a0297360f6c04e90f30df2459 QUANTIZATION.md
18
  065a412a4373a0935c700638884aafc54aff2424e5096a8e05a1a3b1924f8071 README.ja.md
19
  c44fa06d211d8619913152504c588d9ab955bf09eb04f547a23ac4b220143ef3 README.ko.md
20
  1961f17a969707f1c99ccaf8f66c598fd24a53087ffccb173653bbdb60af4671 README.md
@@ -65,13 +65,13 @@ ce42807abfea596c63f1cc6524204de2ec242d12305569373ec7ee6cb5ea8963 configuration_
65
  a8e096a98853840e8838c951ff887c34c751ae137ba926597561e61ca8de96ef evaluation/QEV_DLP_유형별_성능평가_종합보고서.md
66
  7f23b55cf7a3058b334bcda6717b3655d45676606da90864ab92447b91637cc5 evaluation/QEV_DLP_유형별_성능평가_종합보고서.pdf
67
  4a6496521ec5ee23e66e05d37af9940150454907a6dfca698d23fae7169e1d12 evaluation/qev-type-eval-actual-int8-summary.json
68
- 8398ab7e9090c698c0f2ed26f38fef74e02a01870b4e848665f3b056fb6f6272 evaluation.json
69
  986fb901f7edb271d343641dd3d0091c9bfdb82525ebc027ca41dc4e54162e8c example.py
70
  b8b6ed5e9af2029dde9f99147456242064ce37f9ca1cc46c1fab8231d52b325e head_qev.py
71
  da8ada934aaef29b39fd5d65ff7184e62f467c17653156b8260316494b9443a0 modeling_qev.py
72
  45256d0fd69fbebda3ba177d8c9d7a1ffb0cc66577033e5183649e6b81e8b94e modeling_qev_public.py
73
  4a368afaadb2fef63dd755a2c6759dd470b45213e167228aa4ad3b9ae94d0f92 pointer_head.safetensors
74
- a8c0a6bb1ed64fd672bf9d773a5d45c357e18292a325bb6db414488a24370b3a provenance.json
75
  c1e6c40413a71a21c854d814ed6072ca807076c9b0dc511a2a8d401c4335fb56 quantization/config.json
76
  68d84c785586ba5f5150e0cf46a911967764b231839a62240b0f1cf293c515b9 quantization/quantize_int8.py
77
  1807f5729c6813856d4b662bf8ba7c3099543203fc9767767567915374fa8f1f quantization/requirements.txt
 
10
  8d73ef0374c47be1033c1e78cc1427af2cce62c263da3dc06bdb06512d38e2d6 LICENSES/._Apache-2.0.txt
11
  cfc7749b96f63bd31c3c42b5c471bf756814053e847c10f3eb003417bc523d30 LICENSES/Apache-2.0.txt
12
  4b39d0a3e6ff27421e95e7b67378331cc88dd7e2014eb4175f8a5f1c6ee73198 LICENSE_POLICY.md
13
+ 028e4c2206bdf357bea327fa1a0d6fea3f7e6df7f6d3834763d57271f8424cca MODEL_LINEAGE.json
14
  baceafe0782b0655c6147f33d369152bbd084d7d3c372afd4c484384485c029d NOTICE
15
+ 4492921a4fd283e1a30bf591060deab2f2475fc07b8fdc112d8f14685b1c992b QUANTIZATION.ja.md
16
+ 048696be780920beeff9abb5727ed070be685bc2574d196ba8e5b8a2a9216129 QUANTIZATION.ko.md
17
+ eef2d7d8be2816e1fa09864394d4b801224f8755fbc44c901a1c48e197f907ff QUANTIZATION.md
18
  065a412a4373a0935c700638884aafc54aff2424e5096a8e05a1a3b1924f8071 README.ja.md
19
  c44fa06d211d8619913152504c588d9ab955bf09eb04f547a23ac4b220143ef3 README.ko.md
20
  1961f17a969707f1c99ccaf8f66c598fd24a53087ffccb173653bbdb60af4671 README.md
 
65
  a8e096a98853840e8838c951ff887c34c751ae137ba926597561e61ca8de96ef evaluation/QEV_DLP_유형별_성능평가_종합보고서.md
66
  7f23b55cf7a3058b334bcda6717b3655d45676606da90864ab92447b91637cc5 evaluation/QEV_DLP_유형별_성능평가_종합보고서.pdf
67
  4a6496521ec5ee23e66e05d37af9940150454907a6dfca698d23fae7169e1d12 evaluation/qev-type-eval-actual-int8-summary.json
68
+ e986c0defc6489d83085592f8d34cdf880af68e08a6e50906b939a32005aed79 evaluation.json
69
  986fb901f7edb271d343641dd3d0091c9bfdb82525ebc027ca41dc4e54162e8c example.py
70
  b8b6ed5e9af2029dde9f99147456242064ce37f9ca1cc46c1fab8231d52b325e head_qev.py
71
  da8ada934aaef29b39fd5d65ff7184e62f467c17653156b8260316494b9443a0 modeling_qev.py
72
  45256d0fd69fbebda3ba177d8c9d7a1ffb0cc66577033e5183649e6b81e8b94e modeling_qev_public.py
73
  4a368afaadb2fef63dd755a2c6759dd470b45213e167228aa4ad3b9ae94d0f92 pointer_head.safetensors
74
+ 4798a8a6fd7b6497536acfc5a0df7347540d7bed607d82f1ab5df84946014d5c provenance.json
75
  c1e6c40413a71a21c854d814ed6072ca807076c9b0dc511a2a8d401c4335fb56 quantization/config.json
76
  68d84c785586ba5f5150e0cf46a911967764b231839a62240b0f1cf293c515b9 quantization/quantize_int8.py
77
  1807f5729c6813856d4b662bf8ba7c3099543203fc9767767567915374fa8f1f quantization/requirements.txt
MODEL_LINEAGE.json CHANGED
@@ -1,10 +1,10 @@
1
  {
2
- "release_name": "QEV2.1-12B-Dev",
3
  "previous_run_name": "QEV2-Multitask-JA-R1",
4
  "precision": "bf16",
5
  "family": "qev2",
6
- "base_release": "querypieai/QEV2.1-12B-Dev",
7
- "selected_run": "qev2-multitask-ja-20261024-r2/runs/multitask-r1",
8
  "training_data_distributed": false,
9
  "raw_evaluation_rows_distributed": false,
10
  "adapter_and_pointer_head_replaced": true,
 
1
  {
2
+ "release_name": "QEV2.1-12B",
3
  "previous_run_name": "QEV2-Multitask-JA-R1",
4
  "precision": "bf16",
5
  "family": "qev2",
6
+ "base_release": "QEV2.1-12B",
7
+ "selected_run": "public-multitask-release",
8
  "training_data_distributed": false,
9
  "raw_evaluation_rows_distributed": false,
10
  "adapter_and_pointer_head_replaced": true,
QUANTIZATION.ja.md CHANGED
@@ -1,4 +1,4 @@
1
- # QEV2.1-12B-Dev — 8bit量子化ガイド
2
 
3
  [English](QUANTIZATION.md) · [한국어](QUANTIZATION.ko.md) · [日本語](QUANTIZATION.ja.md)
4
 
@@ -10,19 +10,19 @@ CUDA GPUが必要で、H200で検証しています。Mac/MLX・CPU・複数GPU
10
 
11
  ```bash
12
  hf auth login
13
- hf download querypieai/QEV2.1-12B-Dev --local-dir ./QEV2.1-12B-Dev
14
- cd QEV2.1-12B-Dev
15
  pip install -r requirements.txt
16
  pip install -r quantization/requirements.txt
17
 
18
  python quantization/quantize_int8.py --model . \
19
- --out ../QEV2.1-12B-Dev-INT8 --device cuda:0 --local-files-only
20
  ```
21
 
22
  --modelはローカルフォルダーまたはHub repo IDです。Hubは--revisionでcommitを固定できます。private認証はhf auth loginまたはHF_TOKENを使い、コード/ログへtokenを記録しないでください。--local-files-onlyはローカル/キャッシュのみ。--outは原本外の未作成パスが必要で上書きしません。config.jsonはレシピ説明で、CLIの固定検証レシピを実行します。
23
 
24
  ```text
25
- QEV2.1-12B-Dev-INT8/
26
  backbone/ # Serialized INT8 weights + quantization_config
27
  adapter/ # Original unmerged trained adapter
28
  pointer_head.safetensors # Original FP32 pointer head
@@ -39,7 +39,7 @@ Linear8bitLt、原重みとadapter/head/tokenizerのhash不変、確率合計、
39
 
40
  ```python
41
  from transformers import AutoModel
42
- model = AutoModel.from_pretrained("../QEV2.1-12B-Dev-INT8",
43
  trust_remote_code=True, local_files_only=True, device="cuda:0")
44
  result = model.decide("This is a public announcement.", {
45
  "dlp": {"type":"noul", "instructions":"Does this contain personal contact information?",
 
1
+ # QEV2.1-12B — 8bit量子化ガイド
2
 
3
  [English](QUANTIZATION.md) · [한국어](QUANTIZATION.ko.md) · [日本語](QUANTIZATION.ja.md)
4
 
 
10
 
11
  ```bash
12
  hf auth login
13
+ hf download querypieai/QEV2.1-12B --local-dir ./QEV2.1-12B
14
+ cd QEV2.1-12B
15
  pip install -r requirements.txt
16
  pip install -r quantization/requirements.txt
17
 
18
  python quantization/quantize_int8.py --model . \
19
+ --out ../QEV2.1-12B-8BIT --device cuda:0 --local-files-only
20
  ```
21
 
22
  --modelはローカルフォルダーまたはHub repo IDです。Hubは--revisionでcommitを固定できます。private認証はhf auth loginまたはHF_TOKENを使い、コード/ログへtokenを記録しないでください。--local-files-onlyはローカル/キャッシュのみ。--outは原本外の未作成パスが必要で上書きしません。config.jsonはレシピ説明で、CLIの固定検証レシピを実行します。
23
 
24
  ```text
25
+ QEV2.1-12B-8BIT/
26
  backbone/ # Serialized INT8 weights + quantization_config
27
  adapter/ # Original unmerged trained adapter
28
  pointer_head.safetensors # Original FP32 pointer head
 
39
 
40
  ```python
41
  from transformers import AutoModel
42
+ model = AutoModel.from_pretrained("../QEV2.1-12B-8BIT",
43
  trust_remote_code=True, local_files_only=True, device="cuda:0")
44
  result = model.decide("This is a public announcement.", {
45
  "dlp": {"type":"noul", "instructions":"Does this contain personal contact information?",
QUANTIZATION.ko.md CHANGED
@@ -1,4 +1,4 @@
1
- # QEV2.1-12B-Dev — 8bit 양자화 가이드
2
 
3
  [English](QUANTIZATION.md) · [한국어](QUANTIZATION.ko.md) · [日本語](QUANTIZATION.ja.md)
4
 
@@ -10,19 +10,19 @@ CUDA GPU가 필요합니다. 이 도구는 H200에서 검증했으며 Mac/MLX·C
10
 
11
  ```bash
12
  hf auth login
13
- hf download querypieai/QEV2.1-12B-Dev --local-dir ./QEV2.1-12B-Dev
14
- cd QEV2.1-12B-Dev
15
  pip install -r requirements.txt
16
  pip install -r quantization/requirements.txt
17
 
18
  python quantization/quantize_int8.py --model . \
19
- --out ../QEV2.1-12B-Dev-INT8 --device cuda:0 --local-files-only
20
  ```
21
 
22
  --model은 로컬 폴더 또는 Hub repo ID입니다. Hub repo ID를 지정할 때는 --revision으로 커밋을 고정할 수 있습니다. 비공개 저장소 접근은 hf auth login 또는 HF_TOKEN 환경변수를 사용하고 토큰을 코드나 로그에 넣지 마세요. --local-files-only는 로컬·캐시만 사용합니다. --out은 존재하지 않는 별도 경로여야 하며 덮어쓰지 않습니다. config.json은 레시피 설명용이며 실제 옵션은 CLI에 정의된 검증 레시피를 사용합니다.
23
 
24
  ```text
25
- QEV2.1-12B-Dev-INT8/
26
  backbone/ # Serialized INT8 weights + quantization_config
27
  adapter/ # Original unmerged trained adapter
28
  pointer_head.safetensors # Original FP32 pointer head
@@ -39,7 +39,7 @@ QEV2.1-12B-Dev-INT8/
39
 
40
  ```python
41
  from transformers import AutoModel
42
- model = AutoModel.from_pretrained("../QEV2.1-12B-Dev-INT8",
43
  trust_remote_code=True, local_files_only=True, device="cuda:0")
44
  result = model.decide("This is a public announcement.", {
45
  "dlp": {"type":"noul", "instructions":"Does this contain personal contact information?",
 
1
+ # QEV2.1-12B — 8bit 양자화 가이드
2
 
3
  [English](QUANTIZATION.md) · [한국어](QUANTIZATION.ko.md) · [日本語](QUANTIZATION.ja.md)
4
 
 
10
 
11
  ```bash
12
  hf auth login
13
+ hf download querypieai/QEV2.1-12B --local-dir ./QEV2.1-12B
14
+ cd QEV2.1-12B
15
  pip install -r requirements.txt
16
  pip install -r quantization/requirements.txt
17
 
18
  python quantization/quantize_int8.py --model . \
19
+ --out ../QEV2.1-12B-8BIT --device cuda:0 --local-files-only
20
  ```
21
 
22
  --model은 로컬 폴더 또는 Hub repo ID입니다. Hub repo ID를 지정할 때는 --revision으로 커밋을 고정할 수 있습니다. 비공개 저장소 접근은 hf auth login 또는 HF_TOKEN 환경변수를 사용하고 토큰을 코드나 로그에 넣지 마세요. --local-files-only는 로컬·캐시만 사용합니다. --out은 존재하지 않는 별도 경로여야 하며 덮어쓰지 않습니다. config.json은 레시피 설명용이며 실제 옵션은 CLI에 정의된 검증 레시피를 사용합니다.
23
 
24
  ```text
25
+ QEV2.1-12B-8BIT/
26
  backbone/ # Serialized INT8 weights + quantization_config
27
  adapter/ # Original unmerged trained adapter
28
  pointer_head.safetensors # Original FP32 pointer head
 
39
 
40
  ```python
41
  from transformers import AutoModel
42
+ model = AutoModel.from_pretrained("../QEV2.1-12B-8BIT",
43
  trust_remote_code=True, local_files_only=True, device="cuda:0")
44
  result = model.decide("This is a public announcement.", {
45
  "dlp": {"type":"noul", "instructions":"Does this contain personal contact information?",
QUANTIZATION.md CHANGED
@@ -1,4 +1,4 @@
1
- # QEV2.1-12B-Dev — 8-bit quantization guide
2
 
3
  [English](QUANTIZATION.md) · [한국어](QUANTIZATION.ko.md) · [日本語](QUANTIZATION.ja.md)
4
 
@@ -10,19 +10,19 @@ CUDA is required. Validated on H200; not a Mac/MLX, CPU or multi-GPU conversion
10
 
11
  ```bash
12
  hf auth login
13
- hf download querypieai/QEV2.1-12B-Dev --local-dir ./QEV2.1-12B-Dev
14
- cd QEV2.1-12B-Dev
15
  pip install -r requirements.txt
16
  pip install -r quantization/requirements.txt
17
 
18
  python quantization/quantize_int8.py --model . \
19
- --out ../QEV2.1-12B-Dev-INT8 --device cuda:0 --local-files-only
20
  ```
21
 
22
  --model accepts a local folder or Hub repo ID; pin a Hub commit with --revision. Authenticate private access using hf auth login or HF_TOKEN, never a token in code/logs. --local-files-only uses local/cache assets only. --out must not exist and must be outside the source. config.json documents the recipe; the CLI runs its validated fixed recipe.
23
 
24
  ```text
25
- QEV2.1-12B-Dev-INT8/
26
  backbone/ # Serialized INT8 weights + quantization_config
27
  adapter/ # Original unmerged trained adapter
28
  pointer_head.safetensors # Original FP32 pointer head
@@ -39,7 +39,7 @@ The tool checks Linear8bitLt modules, unchanged source and adapter/head/tokenize
39
 
40
  ```python
41
  from transformers import AutoModel
42
- model = AutoModel.from_pretrained("../QEV2.1-12B-Dev-INT8",
43
  trust_remote_code=True, local_files_only=True, device="cuda:0")
44
  result = model.decide("This is a public announcement.", {
45
  "dlp": {"type":"noul", "instructions":"Does this contain personal contact information?",
 
1
+ # QEV2.1-12B — 8-bit quantization guide
2
 
3
  [English](QUANTIZATION.md) · [한국어](QUANTIZATION.ko.md) · [日本語](QUANTIZATION.ja.md)
4
 
 
10
 
11
  ```bash
12
  hf auth login
13
+ hf download querypieai/QEV2.1-12B --local-dir ./QEV2.1-12B
14
+ cd QEV2.1-12B
15
  pip install -r requirements.txt
16
  pip install -r quantization/requirements.txt
17
 
18
  python quantization/quantize_int8.py --model . \
19
+ --out ../QEV2.1-12B-8BIT --device cuda:0 --local-files-only
20
  ```
21
 
22
  --model accepts a local folder or Hub repo ID; pin a Hub commit with --revision. Authenticate private access using hf auth login or HF_TOKEN, never a token in code/logs. --local-files-only uses local/cache assets only. --out must not exist and must be outside the source. config.json documents the recipe; the CLI runs its validated fixed recipe.
23
 
24
  ```text
25
+ QEV2.1-12B-8BIT/
26
  backbone/ # Serialized INT8 weights + quantization_config
27
  adapter/ # Original unmerged trained adapter
28
  pointer_head.safetensors # Original FP32 pointer head
 
39
 
40
  ```python
41
  from transformers import AutoModel
42
+ model = AutoModel.from_pretrained("../QEV2.1-12B-8BIT",
43
  trust_remote_code=True, local_files_only=True, device="cuda:0")
44
  result = model.decide("This is a public announcement.", {
45
  "dlp": {"type":"noul", "instructions":"Does this contain personal contact information?",
evaluation.json CHANGED
@@ -1,5 +1,5 @@
1
  {
2
- "model": "QEV2.1-12B-Dev",
3
  "precision": "bf16",
4
  "metric": "direct type-question argmax",
5
  "records": 6281,
@@ -40,6 +40,6 @@
40
  "is_iban": 1.0
41
  },
42
  "biometric_note": "is_biometric has 0 positive rows in this split and is excluded from macro summaries.",
43
- "source_report": "qev-type-eval-actual-int8-summary.json",
44
  "scope": "Historical unmerged BF16 source only; NOT newly measured merged/INT8 benchmark scores"
45
  }
 
1
  {
2
+ "model": "QEV2.1-12B",
3
  "precision": "bf16",
4
  "metric": "direct type-question argmax",
5
  "records": 6281,
 
40
  "is_iban": 1.0
41
  },
42
  "biometric_note": "is_biometric has 0 positive rows in this split and is excluded from macro summaries.",
43
+ "source_report": "evaluation/qev-type-eval-actual-int8-summary.json",
44
  "scope": "Historical unmerged BF16 source only; NOT newly measured merged/INT8 benchmark scores"
45
  }
provenance.json CHANGED
@@ -1,11 +1,11 @@
1
  {
2
- "release_name": "QEV2.1-12B-Dev",
3
  "previous_run_name": "QEV2-Multitask-JA-R1",
4
  "precision": "bf16",
5
  "family": "qev2",
6
  "base_model": "google/gemma-4-12B-it",
7
  "base_model_revision": "707f0a3b8a3c7ad586ed01e27eafbad8a27dd0f7",
8
- "selected_checkpoint": "/home/irteam/qev2-multitask-ja-20261024-r2/runs/multitask-r1/checkpoint",
9
  "training_data_distributed": false,
10
  "raw_evaluation_rows_distributed": false,
11
  "self_contained_backbone": true,
@@ -13,9 +13,7 @@
13
  "pointer_head_format": "safetensors",
14
  "distribution": "inference-only",
15
  "source_weights_sha256": {
16
- "pointer_head.safetensors": "4a368afaadb2fef63dd755a2c6759dd470b45213e167228aa4ad3b9ae94d0f92",
17
- "backbone/model.safetensors": "5a84cb313260ac447237b890387116dfa8682e49a6b44bc585ae8353abbff18d",
18
- "adapter/adapter_model.safetensors": "7b9fc3019e3f92eee73924eaf17f1064405cca07437017032b7df7052c08787c"
19
  },
20
  "upload_performed": false
21
  }
 
1
  {
2
+ "release_name": "QEV2.1-12B",
3
  "previous_run_name": "QEV2-Multitask-JA-R1",
4
  "precision": "bf16",
5
  "family": "qev2",
6
  "base_model": "google/gemma-4-12B-it",
7
  "base_model_revision": "707f0a3b8a3c7ad586ed01e27eafbad8a27dd0f7",
8
+ "selected_checkpoint": "public-release-artifact",
9
  "training_data_distributed": false,
10
  "raw_evaluation_rows_distributed": false,
11
  "self_contained_backbone": true,
 
13
  "pointer_head_format": "safetensors",
14
  "distribution": "inference-only",
15
  "source_weights_sha256": {
16
+ "release_bundle": "see MANIFEST.sha256"
 
 
17
  },
18
  "upload_performed": false
19
  }