macontreras98 commited on
Commit
8e9c5b9
·
verified ·
1 Parent(s): b588b08

Upload folder using huggingface_hub

Browse files
.gitattributes CHANGED
@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
33
  *.zip filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
 
 
33
  *.zip filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
36
+ tokenizer.json filter=lfs diff=lfs merge=lfs -text
README.md CHANGED
@@ -1,3 +1,117 @@
1
  ---
 
 
 
 
2
  license: apache-2.0
 
 
 
 
 
 
 
 
3
  ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
2
+ base_model: google/gemma-4-31B-it
3
+ base_model_relation: finetune
4
+ library_name: transformers
5
+ pipeline_tag: image-text-to-text
6
  license: apache-2.0
7
+ tags:
8
+ - gemma
9
+ - gemma-4
10
+ - clinical
11
+ - medical
12
+ - clinical-reasoning
13
+ - icu
14
+ - clin-react
15
  ---
16
+
17
+ # Clin-REACT-31B
18
+
19
+ > **Model name in the manuscript:** Clin-REACT 31B
20
+ > **Base model:** `google/gemma-4-31B-it`
21
+ > **Release format:** merged full-parameter checkpoint
22
+
23
+ Clin-REACT is a family of large language models fine-tuned for clinical reasoning using the ICU-REACT framework. The models are designed to reason over clinically relevant information and produce responses for tasks spanning ICU decision-making and broader clinical-reasoning benchmarks.
24
+
25
+ This repository contains the **merged, self-contained checkpoint**. The LoRA adapter used during supervised fine-tuning has been merged into the corresponding base-model weights for distribution and inference.
26
+
27
+ The underlying Gemma 4 architecture is multimodal. The Clin-REACT fine-tuning and benchmark results reported in this model card are **text-based**; inherited multimodal capabilities were not evaluated as part of these results.
28
+
29
+ ## Benchmark results
30
+
31
+ Clin-REACT 31B is compared with its Gemma 4 31B IT backbone and selected strong open models, including substantially larger baselines.
32
+
33
+ Scores below are the **mean primary benchmark scores**, reported as percentages. Higher is better. The macro average is the unweighted mean of the five benchmark primary scores. **Bold** indicates the best result within the comparison set shown for each benchmark.
34
+
35
+ The benchmarks contain 71 ICU-REACT cases, 174 SCT-Bench cases, 72 ER-Reason cases, 1,254 MedRBench cases, and 934 VivaBench cases. Because the benchmarks use different task formulations and primary scoring procedures, individual benchmark scores should primarily be interpreted within each benchmark.
36
+
37
+ | Model | ICU-REACT (n=71) | SCT-Bench (n=174) | ER-Reason (n=72) | MedRBench (n=1254) | VivaBench (n=934) | Macro avg. |
38
+ |---|---:|---:|---:|---:|---:|---:|
39
+ | Clin-REACT 31B | **44.8** | 75.5 | **51.4** | 46.9 | **33.5** | **50.4** |
40
+ | Gemma 4 31B IT | 40.6 | **77.6** | 47.4 | 44.0 | 31.6 | 48.3 |
41
+ | Baichuan M2 32B | 36.9 | 63.0 | 46.2 | 38.4 | 26.9 | 42.3 |
42
+ | Llama 3.3 70B Instruct | 30.9 | 59.9 | 45.2 | 45.1 | 26.7 | 41.6 |
43
+ | GPT-OSS-120B | 40.1 | 74.3 | 48.3 | **47.0** | 32.5 | 48.5 |
44
+
45
+ ## Installation
46
+
47
+ ```bash
48
+ pip install -U torch transformers accelerate
49
+ ```
50
+
51
+ ## Load and use the model
52
+
53
+ Gemma 4 uses a multimodal processor/model interface even for text-only prompts.
54
+
55
+ ```python
56
+ import torch
57
+ from transformers import AutoProcessor, AutoModelForMultimodalLM
58
+
59
+ MODEL_ID = "macontreras98/Clin-REACT-31B"
60
+
61
+ processor = AutoProcessor.from_pretrained(MODEL_ID)
62
+ model = AutoModelForMultimodalLM.from_pretrained(
63
+ MODEL_ID,
64
+ dtype="auto",
65
+ device_map="auto",
66
+ )
67
+
68
+ messages = [
69
+ {"role": "system", "content": "You are a helpful clinical reasoning assistant."},
70
+ {"role": "user", "content": "Enter your clinical question or case here."},
71
+ ]
72
+
73
+ inputs = processor.apply_chat_template(
74
+ messages,
75
+ tokenize=True,
76
+ return_dict=True,
77
+ return_tensors="pt",
78
+ add_generation_prompt=True,
79
+ ).to(model.device)
80
+
81
+ input_len = inputs["input_ids"].shape[-1]
82
+
83
+ with torch.inference_mode():
84
+ outputs = model.generate(
85
+ **inputs,
86
+ max_new_tokens=512,
87
+ do_sample=False,
88
+ )
89
+
90
+ response = processor.decode(
91
+ outputs[0][input_len:],
92
+ skip_special_tokens=False,
93
+ )
94
+
95
+ # Gemma 4 provides a helper for parsing the generated response.
96
+ parsed = processor.parse_response(
97
+ response,
98
+ prefix=inputs["input_ids"],
99
+ )
100
+ print(parsed)
101
+ ```
102
+
103
+ For exact benchmark reproduction, use the same prompting, processor/chat template, preprocessing, thinking-mode configuration, and decoding settings used in the ICU-REACT evaluation pipeline.
104
+
105
+ ## Intended use and limitations
106
+
107
+ Clin-REACT is released for research on clinical reasoning and medical AI. It is **not a medical device** and should not be used as a substitute for professional clinical judgment, diagnosis, treatment decisions, or other autonomous patient-care decisions.
108
+
109
+ Benchmark performance does not establish clinical safety, prospective effectiveness, or suitability for deployment. Users are responsible for evaluating the model for their own setting and for complying with applicable privacy, security, institutional, and regulatory requirements.
110
+
111
+ ## License
112
+
113
+ The base model `google/gemma-4-31B-it` is licensed under the **Apache License 2.0**. This repository distributes a modified derivative checkpoint. See the `LICENSE` file and retain any applicable upstream copyright and attribution notices.
114
+
115
+ ## Citation
116
+
117
+ If you use Clin-REACT in your research, please cite the associated ICU-REACT / Clin-REACT manuscript. A finalized BibTeX citation can be added here upon publication.
chat_template.jinja ADDED
@@ -0,0 +1,266 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {%- macro format_parameters(properties, required) -%}
2
+ {%- set standard_keys = ['description', 'type', 'properties', 'required', 'nullable'] -%}
3
+ {%- set ns = namespace(found_first=false) -%}
4
+ {%- for key, value in properties | dictsort -%}
5
+ {%- set add_comma = false -%}
6
+ {%- if key not in standard_keys -%}
7
+ {%- if ns.found_first %},{% endif -%}
8
+ {%- set ns.found_first = true -%}
9
+ {{ key }}:{
10
+ {%- if value['description'] -%}
11
+ description:<|"|>{{ value['description'] }}<|"|>
12
+ {%- set add_comma = true -%}
13
+ {%- endif -%}
14
+ {%- if value['nullable'] %}
15
+ {%- if add_comma %},{%- else -%} {%- set add_comma = true -%} {% endif -%}
16
+ nullable:true
17
+ {%- endif -%}
18
+ {%- if value['type'] | upper == 'STRING' -%}
19
+ {%- if value['enum'] -%}
20
+ {%- if add_comma %},{%- else -%} {%- set add_comma = true -%} {% endif -%}
21
+ enum:{{ format_argument(value['enum']) }}
22
+ {%- endif -%}
23
+ {%- elif value['type'] | upper == 'OBJECT' -%}
24
+ ,properties:{
25
+ {%- if value['properties'] is defined and value['properties'] is mapping -%}
26
+ {{- format_parameters(value['properties'], value['required'] | default([])) -}}
27
+ {%- elif value is mapping -%}
28
+ {{- format_parameters(value, value['required'] | default([])) -}}
29
+ {%- endif -%}
30
+ }
31
+ {%- if value['required'] -%}
32
+ ,required:[
33
+ {%- for item in value['required'] | default([]) -%}
34
+ <|"|>{{- item -}}<|"|>
35
+ {%- if not loop.last %},{% endif -%}
36
+ {%- endfor -%}
37
+ ]
38
+ {%- endif -%}
39
+ {%- elif value['type'] | upper == 'ARRAY' -%}
40
+ {%- if value['items'] is mapping and value['items'] -%}
41
+ ,items:{
42
+ {%- set ns_items = namespace(found_first=false) -%}
43
+ {%- for item_key, item_value in value['items'] | dictsort -%}
44
+ {%- if item_value is not none -%}
45
+ {%- if ns_items.found_first %},{% endif -%}
46
+ {%- set ns_items.found_first = true -%}
47
+ {%- if item_key == 'properties' -%}
48
+ properties:{
49
+ {%- if item_value is mapping -%}
50
+ {{- format_parameters(item_value, value['items']['required'] | default([])) -}}
51
+ {%- endif -%}
52
+ }
53
+ {%- elif item_key == 'required' -%}
54
+ required:[
55
+ {%- for req_item in item_value -%}
56
+ <|"|>{{- req_item -}}<|"|>
57
+ {%- if not loop.last %},{% endif -%}
58
+ {%- endfor -%}
59
+ ]
60
+ {%- elif item_key == 'type' -%}
61
+ {%- if item_value is string -%}
62
+ type:{{ format_argument(item_value | upper) }}
63
+ {%- else -%}
64
+ type:{{ format_argument(item_value | map('upper') | list) }}
65
+ {%- endif -%}
66
+ {%- else -%}
67
+ {{ item_key }}:{{ format_argument(item_value) }}
68
+ {%- endif -%}
69
+ {%- endif -%}
70
+ {%- endfor -%}
71
+ }
72
+ {%- endif -%}
73
+ {%- endif -%}
74
+ {%- if add_comma %},{%- else -%} {%- set add_comma = true -%} {% endif -%}
75
+ type:<|"|>{{ value['type'] | upper }}<|"|>}
76
+ {%- endif -%}
77
+ {%- endfor -%}
78
+ {%- endmacro -%}
79
+ {%- macro format_function_declaration(tool_data) -%}
80
+ declaration:{{- tool_data['function']['name'] -}}{description:<|"|>{{- tool_data['function']['description'] -}}<|"|>
81
+ {%- set params = tool_data['function']['parameters'] -%}
82
+ {%- if params -%}
83
+ ,parameters:{
84
+ {%- if params['properties'] -%}
85
+ properties:{ {{- format_parameters(params['properties'], params['required']) -}} },
86
+ {%- endif -%}
87
+ {%- if params['required'] -%}
88
+ required:[
89
+ {%- for item in params['required'] -%}
90
+ <|"|>{{- item -}}<|"|>
91
+ {{- ',' if not loop.last -}}
92
+ {%- endfor -%}
93
+ ],
94
+ {%- endif -%}
95
+ {%- if params['type'] -%}
96
+ type:<|"|>{{- params['type'] | upper -}}<|"|>}
97
+ {%- endif -%}
98
+ {%- endif -%}
99
+ {%- if 'response' in tool_data['function'] -%}
100
+ {%- set response_declaration = tool_data['function']['response'] -%}
101
+ ,response:{
102
+ {%- if response_declaration['description'] -%}
103
+ description:<|"|>{{- response_declaration['description'] -}}<|"|>,
104
+ {%- endif -%}
105
+ {%- if response_declaration['type'] | upper == 'OBJECT' -%}
106
+ type:<|"|>{{- response_declaration['type'] | upper -}}<|"|>}
107
+ {%- endif -%}
108
+ {%- endif -%}
109
+ }
110
+ {%- endmacro -%}
111
+ {%- macro format_argument(argument, escape_keys=True) -%}
112
+ {%- if argument is string -%}
113
+ {{- '<|"|>' + argument + '<|"|>' -}}
114
+ {%- elif argument is boolean -%}
115
+ {{- 'true' if argument else 'false' -}}
116
+ {%- elif argument is mapping -%}
117
+ {{- '{' -}}
118
+ {%- set ns = namespace(found_first=false) -%}
119
+ {%- for key, value in argument | dictsort -%}
120
+ {%- if ns.found_first %},{% endif -%}
121
+ {%- set ns.found_first = true -%}
122
+ {%- if escape_keys -%}
123
+ {{- '<|"|>' + key + '<|"|>' -}}
124
+ {%- else -%}
125
+ {{- key -}}
126
+ {%- endif -%}
127
+ :{{- format_argument(value, escape_keys=escape_keys) -}}
128
+ {%- endfor -%}
129
+ {{- '}' -}}
130
+ {%- elif argument is sequence -%}
131
+ {{- '[' -}}
132
+ {%- for item in argument -%}
133
+ {{- format_argument(item, escape_keys=escape_keys) -}}
134
+ {%- if not loop.last %},{% endif -%}
135
+ {%- endfor -%}
136
+ {{- ']' -}}
137
+ {%- else -%}
138
+ {{- argument -}}
139
+ {%- endif -%}
140
+ {%- endmacro -%}
141
+ {%- macro strip_thinking(text) -%}
142
+ {%- set ns = namespace(result='') -%}
143
+ {%- for part in text.split('<channel|>') -%}
144
+ {%- if '<|channel>' in part -%}
145
+ {%- set ns.result = ns.result + part.split('<|channel>')[0] -%}
146
+ {%- else -%}
147
+ {%- set ns.result = ns.result + part -%}
148
+ {%- endif -%}
149
+ {%- endfor -%}
150
+ {{- ns.result | trim -}}
151
+ {%- endmacro -%}
152
+
153
+ {%- set ns = namespace(prev_message_type=None) -%}
154
+ {%- set loop_messages = messages -%}
155
+ {{ bos_token }}
156
+ {#- Handle System/Tool Definitions Block -#}
157
+ {%- if (enable_thinking is defined and enable_thinking) or tools or messages[0]['role'] in ['system', 'developer'] -%}
158
+ {{- '<|turn>system\n' -}}
159
+
160
+ {#- Inject Thinking token at the very top of the FIRST system turn -#}
161
+ {%- if enable_thinking is defined and enable_thinking -%}
162
+ {{- '<|think|>' -}}
163
+ {%- set ns.prev_message_type = 'think' -%}
164
+ {%- endif -%}
165
+
166
+ {%- if messages[0]['role'] in ['system', 'developer'] -%}
167
+ {{- messages[0]['content'] | trim -}}
168
+ {%- set loop_messages = messages[1:] -%}
169
+ {%- endif -%}
170
+
171
+ {%- if tools -%}
172
+ {%- for tool in tools %}
173
+ {{- '<|tool>' -}}
174
+ {{- format_function_declaration(tool) | trim -}}
175
+ {{- '<tool|>' -}}
176
+ {%- endfor %}
177
+ {%- set ns.prev_message_type = 'tool' -%}
178
+ {%- endif -%}
179
+
180
+ {{- '<turn|>\n' -}}
181
+ {%- endif %}
182
+
183
+ {#- Loop through messages -#}
184
+ {%- for message in loop_messages -%}
185
+ {%- set ns.prev_message_type = None -%}
186
+ {%- set role = 'model' if message['role'] == 'assistant' else message['role'] -%}
187
+ {{- '<|turn>' + role + '\n' }}
188
+
189
+ {%- if message['tool_calls'] -%}
190
+ {%- for tool_call in message['tool_calls'] -%}
191
+ {%- set function = tool_call['function'] -%}
192
+ {{- '<|tool_call>call:' + function['name'] + '{' -}}
193
+ {%- if function['arguments'] is mapping -%}
194
+ {%- set ns_args = namespace(found_first=false) -%}
195
+ {%- for key, value in function['arguments'] | dictsort -%}
196
+ {%- if ns_args.found_first %},{% endif -%}
197
+ {%- set ns_args.found_first = true -%}
198
+ {{- key -}}:{{- format_argument(value, escape_keys=False) -}}
199
+ {%- endfor -%}
200
+ {%- elif function['arguments'] is string -%}
201
+ {{- function['arguments'] -}}
202
+ {%- endif -%}
203
+ {{- '}<tool_call|>' -}}
204
+ {%- endfor -%}
205
+ {%- set ns.prev_message_type = 'tool_call' -%}
206
+ {%- endif -%}
207
+
208
+ {%- if message['tool_responses'] -%}
209
+ {#- Tool Response handling -#}
210
+ {%- for tool_response in message['tool_responses'] -%}
211
+ {{- '<|tool_response>' -}}
212
+ {%- if tool_response['response'] is mapping -%}
213
+ {{- 'response:' + tool_response['name'] | default('unknown') + '{' -}}
214
+ {%- for key, value in tool_response['response'] | dictsort -%}
215
+ {{- key -}}:{{- format_argument(value, escape_keys=False) -}}
216
+ {%- if not loop.last %},{% endif -%}
217
+ {%- endfor -%}
218
+ {{- '}' -}}
219
+ {%- else -%}
220
+ {{- 'response:' + tool_response['name'] | default('unknown') + '{value:' + format_argument(tool_response['response'], escape_keys=False) + '}' -}}
221
+ {%- endif -%}
222
+ {{- '<tool_response|>' -}}
223
+ {%- endfor -%}
224
+ {%- set ns.prev_message_type = 'tool_response' -%}
225
+ {%- endif -%}
226
+
227
+ {%- if message['content'] is string -%}
228
+ {%- if role == 'model' -%}
229
+ {{- strip_thinking(message['content']) -}}
230
+ {%- else -%}
231
+ {{- message['content'] | trim -}}
232
+ {%- endif -%}
233
+ {%- elif message['content'] is sequence -%}
234
+ {%- for item in message['content'] -%}
235
+ {%- if item['type'] == 'text' -%}
236
+ {%- if role == 'model' -%}
237
+ {{- strip_thinking(item['text']) -}}
238
+ {%- else -%}
239
+ {{- item['text'] | trim -}}
240
+ {%- endif -%}
241
+ {%- elif item['type'] == 'image' -%}
242
+ {{- '\n\n<|image|>\n\n' -}}
243
+ {%- set ns.prev_message_type = 'image' -%}
244
+ {%- elif item['type'] == 'audio' -%}
245
+ {{- '<|audio|>' -}}
246
+ {%- set ns.prev_message_type = 'audio' -%}
247
+ {%- elif item['type'] == 'video' -%}
248
+ {{- '\n\n<|video|>\n\n' -}}
249
+ {%- set ns.prev_message_type = 'video' -%}
250
+ {%- endif -%}
251
+ {%- endfor -%}
252
+ {%- endif -%}
253
+
254
+ {%- if not (message['tool_responses'] and not message['content']) -%}
255
+ {{- '<turn|>\n' -}}
256
+ {%- endif -%}
257
+ {%- endfor -%}
258
+
259
+ {%- if add_generation_prompt -%}
260
+ {%- if ns.prev_message_type != 'tool_response' -%}
261
+ {{- '<|turn>model\n' -}}
262
+ {%- endif -%}
263
+ {%- if not enable_thinking | default(false) -%}
264
+ {{- '<|channel>thought\n<channel|>' -}}
265
+ {%- endif -%}
266
+ {%- endif -%}
config.json ADDED
@@ -0,0 +1,177 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architectures": [
3
+ "Gemma4ForConditionalGeneration"
4
+ ],
5
+ "audio_config": null,
6
+ "audio_token_id": 258881,
7
+ "boa_token_id": 256000,
8
+ "boi_token_id": 255999,
9
+ "dtype": "bfloat16",
10
+ "eoa_token_id": 258883,
11
+ "eoa_token_index": 258883,
12
+ "eoi_token_id": 258882,
13
+ "eos_token_id": [
14
+ 1,
15
+ 106
16
+ ],
17
+ "image_token_id": 258880,
18
+ "initializer_range": 0.02,
19
+ "model_type": "gemma4",
20
+ "text_config": {
21
+ "attention_bias": false,
22
+ "attention_dropout": 0.0,
23
+ "attention_k_eq_v": true,
24
+ "bos_token_id": 2,
25
+ "dtype": "bfloat16",
26
+ "enable_moe_block": false,
27
+ "eos_token_id": 1,
28
+ "expert_intermediate_size": null,
29
+ "final_logit_softcapping": 30.0,
30
+ "global_head_dim": 512,
31
+ "head_dim": 256,
32
+ "hidden_activation": "gelu_pytorch_tanh",
33
+ "hidden_size": 5376,
34
+ "hidden_size_per_layer_input": 0,
35
+ "initializer_range": 0.02,
36
+ "intermediate_size": 21504,
37
+ "layer_types": [
38
+ "sliding_attention",
39
+ "sliding_attention",
40
+ "sliding_attention",
41
+ "sliding_attention",
42
+ "sliding_attention",
43
+ "full_attention",
44
+ "sliding_attention",
45
+ "sliding_attention",
46
+ "sliding_attention",
47
+ "sliding_attention",
48
+ "sliding_attention",
49
+ "full_attention",
50
+ "sliding_attention",
51
+ "sliding_attention",
52
+ "sliding_attention",
53
+ "sliding_attention",
54
+ "sliding_attention",
55
+ "full_attention",
56
+ "sliding_attention",
57
+ "sliding_attention",
58
+ "sliding_attention",
59
+ "sliding_attention",
60
+ "sliding_attention",
61
+ "full_attention",
62
+ "sliding_attention",
63
+ "sliding_attention",
64
+ "sliding_attention",
65
+ "sliding_attention",
66
+ "sliding_attention",
67
+ "full_attention",
68
+ "sliding_attention",
69
+ "sliding_attention",
70
+ "sliding_attention",
71
+ "sliding_attention",
72
+ "sliding_attention",
73
+ "full_attention",
74
+ "sliding_attention",
75
+ "sliding_attention",
76
+ "sliding_attention",
77
+ "sliding_attention",
78
+ "sliding_attention",
79
+ "full_attention",
80
+ "sliding_attention",
81
+ "sliding_attention",
82
+ "sliding_attention",
83
+ "sliding_attention",
84
+ "sliding_attention",
85
+ "full_attention",
86
+ "sliding_attention",
87
+ "sliding_attention",
88
+ "sliding_attention",
89
+ "sliding_attention",
90
+ "sliding_attention",
91
+ "full_attention",
92
+ "sliding_attention",
93
+ "sliding_attention",
94
+ "sliding_attention",
95
+ "sliding_attention",
96
+ "sliding_attention",
97
+ "full_attention"
98
+ ],
99
+ "max_position_embeddings": 262144,
100
+ "model_type": "gemma4_text",
101
+ "moe_intermediate_size": null,
102
+ "num_attention_heads": 32,
103
+ "num_experts": null,
104
+ "num_global_key_value_heads": 4,
105
+ "num_hidden_layers": 60,
106
+ "num_key_value_heads": 16,
107
+ "num_kv_shared_layers": 0,
108
+ "pad_token_id": 0,
109
+ "rms_norm_eps": 1e-06,
110
+ "rope_parameters": {
111
+ "full_attention": {
112
+ "partial_rotary_factor": 0.25,
113
+ "rope_theta": 1000000.0,
114
+ "rope_type": "proportional"
115
+ },
116
+ "sliding_attention": {
117
+ "rope_theta": 10000.0,
118
+ "rope_type": "default"
119
+ }
120
+ },
121
+ "sliding_window": 1024,
122
+ "tie_word_embeddings": true,
123
+ "top_k_experts": null,
124
+ "use_bidirectional_attention": "vision",
125
+ "use_cache": true,
126
+ "use_double_wide_mlp": false,
127
+ "vocab_size": 262144,
128
+ "vocab_size_per_layer_input": 262144
129
+ },
130
+ "tie_word_embeddings": true,
131
+ "transformers_version": "5.5.0",
132
+ "video_token_id": 258884,
133
+ "vision_config": {
134
+ "_name_or_path": "",
135
+ "architectures": null,
136
+ "attention_bias": false,
137
+ "attention_dropout": 0.0,
138
+ "chunk_size_feed_forward": 0,
139
+ "default_output_length": 280,
140
+ "dtype": "bfloat16",
141
+ "global_head_dim": 72,
142
+ "head_dim": 72,
143
+ "hidden_activation": "gelu_pytorch_tanh",
144
+ "hidden_size": 1152,
145
+ "id2label": {
146
+ "0": "LABEL_0",
147
+ "1": "LABEL_1"
148
+ },
149
+ "initializer_range": 0.02,
150
+ "intermediate_size": 4304,
151
+ "is_encoder_decoder": false,
152
+ "label2id": {
153
+ "LABEL_0": 0,
154
+ "LABEL_1": 1
155
+ },
156
+ "max_position_embeddings": 131072,
157
+ "model_type": "gemma4_vision",
158
+ "num_attention_heads": 16,
159
+ "num_hidden_layers": 27,
160
+ "num_key_value_heads": 16,
161
+ "output_attentions": false,
162
+ "output_hidden_states": false,
163
+ "patch_size": 16,
164
+ "pooling_kernel_size": 3,
165
+ "position_embedding_size": 10240,
166
+ "problem_type": null,
167
+ "return_dict": true,
168
+ "rms_norm_eps": 1e-06,
169
+ "rope_parameters": {
170
+ "rope_theta": 100.0,
171
+ "rope_type": "default"
172
+ },
173
+ "standardize": true,
174
+ "use_clipped_linears": false
175
+ },
176
+ "vision_soft_tokens_per_image": 280
177
+ }
generation_config.json ADDED
@@ -0,0 +1,14 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "bos_token_id": 2,
3
+ "do_sample": true,
4
+ "eos_token_id": [
5
+ 1,
6
+ 106,
7
+ 50
8
+ ],
9
+ "pad_token_id": 0,
10
+ "temperature": 1.0,
11
+ "top_k": 64,
12
+ "top_p": 0.95,
13
+ "transformers_version": "5.5.0"
14
+ }
model-00001-of-00013.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7dd3757d78ec63db96809d24d8fee6f6f7c335a9aea9eb65535c5b659939e521
3
+ size 4978021046
model-00002-of-00013.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5d67d60a21b43b892f70b1a88ebb3266455599fa1d916d3834a4b745c3998f20
3
+ size 4899701506
model-00003-of-00013.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:921e98137b66c4e69f56bb1f329f091d36790b9d5961a8ef38a876938247635f
3
+ size 4899701514
model-00004-of-00013.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:dd71d6842c8139184d843784252beab39c4fc708a9f0bd369dcf59bf0fe678ca
3
+ size 4789600202
model-00005-of-00013.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:352b37908564616773539a5246e36548e33e475bf751a74bb359024a6c9c7ca3
3
+ size 4899701578
model-00006-of-00013.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:80537e9165e1d591cdf232faf21cf6fe4fba15fe79385b18c51f9df1b52e101a
3
+ size 4899701578
model-00007-of-00013.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b8bc2b5cda723305a3c331908a8d224f8b04d63d35b0c70b59fa2edb41acf831
3
+ size 4899701578
model-00008-of-00013.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5cf1936f4f9c808a85993598318cb43805b671e2161a612c7d25fef6f33dd660
3
+ size 4899701578
model-00009-of-00013.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:79a10fcc9a2ba77113399ece2c84dc23b7665fcf0751c0b9c0525ffadf9031e6
3
+ size 4899701578
model-00010-of-00013.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:2200b4811deb363952f7d56e18df163f4421fc65a9da20d8daa9d2cf5a2b691c
3
+ size 4789600202
model-00011-of-00013.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:8f3845fb1d95d12a623c85abc35be862608d78a869dc30e201abc18849013534
3
+ size 4899701578
model-00012-of-00013.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d47c5a70b7e45fcb00b639b71e2144f3b4a30781d1562d6d000dc40569832dac
3
+ size 4899701578
model-00013-of-00013.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0c244fe7d6b0aba5060df626f53182eabd4727650013c94fd29a76b6570a6807
3
+ size 3891800932
model.safetensors.index.json ADDED
The diff for this file is too large to render. See raw diff
 
tokenizer.json ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5d84efaa7efd10a17dd1d85d92456585541d6e6c9ef37eb78b8eb753442905f5
3
+ size 32169725
tokenizer_config.json ADDED
@@ -0,0 +1,99 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "audio_token": "<|audio|>",
3
+ "backend": "tokenizers",
4
+ "boa_token": "<|audio>",
5
+ "boi_token": "<|image>",
6
+ "bos_token": "<bos>",
7
+ "eoa_token": "<audio|>",
8
+ "eoc_token": "<channel|>",
9
+ "eoi_token": "<image|>",
10
+ "eos_token": "<eos>",
11
+ "eot_token": "<turn|>",
12
+ "escape_token": "<|\"|>",
13
+ "etc_token": "<tool_call|>",
14
+ "etd_token": "<tool|>",
15
+ "etr_token": "<tool_response|>",
16
+ "extra_special_tokens": [
17
+ "<|video|>"
18
+ ],
19
+ "image_token": "<|image|>",
20
+ "is_local": true,
21
+ "mask_token": "<mask>",
22
+ "max_length": 2048,
23
+ "model_max_length": 1000000000000000019884624838656,
24
+ "model_specific_special_tokens": {
25
+ "audio_token": "<|audio|>",
26
+ "boa_token": "<|audio>",
27
+ "boi_token": "<|image>",
28
+ "eoa_token": "<audio|>",
29
+ "eoc_token": "<channel|>",
30
+ "eoi_token": "<image|>",
31
+ "eot_token": "<turn|>",
32
+ "escape_token": "<|\"|>",
33
+ "etc_token": "<tool_call|>",
34
+ "etd_token": "<tool|>",
35
+ "etr_token": "<tool_response|>",
36
+ "image_token": "<|image|>",
37
+ "soc_token": "<|channel>",
38
+ "sot_token": "<|turn>",
39
+ "stc_token": "<|tool_call>",
40
+ "std_token": "<|tool>",
41
+ "str_token": "<|tool_response>",
42
+ "think_token": "<|think|>"
43
+ },
44
+ "pad_token": "<pad>",
45
+ "padding_side": "left",
46
+ "processor_class": "Gemma4Processor",
47
+ "response_schema": {
48
+ "properties": {
49
+ "content": {
50
+ "type": "string"
51
+ },
52
+ "role": {
53
+ "const": "assistant"
54
+ },
55
+ "thinking": {
56
+ "type": "string"
57
+ },
58
+ "tool_calls": {
59
+ "items": {
60
+ "properties": {
61
+ "function": {
62
+ "properties": {
63
+ "arguments": {
64
+ "additionalProperties": {},
65
+ "type": "object",
66
+ "x-parser": "gemma4-tool-call"
67
+ },
68
+ "name": {
69
+ "type": "string"
70
+ }
71
+ },
72
+ "type": "object",
73
+ "x-regex": "call\\:(?P<name>\\w+)(?P<arguments>\\{.*\\})"
74
+ },
75
+ "type": {
76
+ "const": "function"
77
+ }
78
+ },
79
+ "type": "object"
80
+ },
81
+ "type": "array",
82
+ "x-regex-iterator": "<\\|tool_call>(.*?)<tool_call\\|>"
83
+ }
84
+ },
85
+ "type": "object",
86
+ "x-regex": "(\\<\\|channel\\>thought\\n(?P<thinking>.*?)\\<channel\\|\\>)?(?P<content>(?:(?!\\<\\|tool_call\\>)(?!\\<turn\\|\\>).)+)?(?P<tool_calls>\\<\\|tool_call\\>.*\\<tool_call\\|\\>)?(?:\\<turn\\|\\>)?"
87
+ },
88
+ "soc_token": "<|channel>",
89
+ "sot_token": "<|turn>",
90
+ "stc_token": "<|tool_call>",
91
+ "std_token": "<|tool>",
92
+ "str_token": "<|tool_response>",
93
+ "stride": 0,
94
+ "think_token": "<|think|>",
95
+ "tokenizer_class": "GemmaTokenizer",
96
+ "truncation_side": "right",
97
+ "truncation_strategy": "longest_first",
98
+ "unk_token": "<unk>"
99
+ }