EdwardIrby commited on
Commit
4adc78d
·
verified ·
1 Parent(s): c3ffb5a

Upload reflex-search v3.3 text-only BF16 + LoRA + calibrated Choice/Noul/Score heads

Browse files
Files changed (8) hide show
  1. .gitattributes +1 -0
  2. README.md +53 -0
  3. calibration.json +104 -0
  4. config.json +208 -0
  5. heads.json +326 -0
  6. heads.safetensors +3 -0
  7. model.safetensors +3 -0
  8. tokenizer.json +3 -0
.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 +1,56 @@
1
  ---
2
  license: apache-2.0
 
 
 
 
 
 
 
 
 
 
 
3
  ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
2
  license: apache-2.0
3
+ base_model: google/embeddinggemma-2
4
+ library_name: safetensors
5
+ pipeline_tag: text-classification
6
+ tags:
7
+ - search
8
+ - reranking
9
+ - calibration
10
+ - system-1
11
+ - cpu-inference
12
+ - embeddinggemma-2
13
+ - mcp
14
  ---
15
+
16
+ # `youdotcom/reflex-search` (`v3.3`)
17
+
18
+ `reflex-search` is a compact, CPU-native System 1 search decision and source-ranking model built on the text encoder of [`google/embeddinggemma-2`](https://huggingface.co/google/embeddinggemma-2) (`d = 768`, 24 text layers with hybrid 512-token sliding and global full attention). Vision and audio towers are removed so the text-only `bf16` checkpoint occupies **517 MB on disk** and runs in **~196 MB RSS (`int8`) / ~315 MB RSS (`bf16`)** on commodity CPUs via `s1_eg2cpu` (adapted from `undrgrnd-labs/hunch/eg2cpu`).
19
+
20
+ It powers the `reflex-search` and `reflex-extract` MCP tools (`/mcp/reflex`):
21
+ 1. **Gate 1 (Sub-query preflight):** Scores parallel candidate sub-queries (`noul_query_matches_intent`, `choice_best_query_for_intent`) to filter off-topic hallucinations and rank over-budget queries before spending Search API calls.
22
+ 2. **Gate 2 (Unified SERP + Knowledge ranking & routing):** Scores web highlights and licensed knowledge facts (`attribution` + `as_of`) on one unified relevance scale (`choice_best_search_hit` + `score_return_relevance`), estimates counterfactual full-page read gain (`noul_hit_worth_reading`), and scores set-level coverage (`score_evidence_coverage` + `choice_next_step` in `{answer, read, revise, stop}`).
23
+ 3. **Parallel Extract Intent Scoring:** Scores BM25-windowed webpage and PDF passages against caller-supplied `document_intent` (`score_extract_relevance` on `doc_kind = "extracted_doc"`).
24
+
25
+ ---
26
+
27
+ ## Architecture & Artifact Layout
28
+
29
+ | File | Format | Size | Contents |
30
+ | :--- | :--- | ---: | :--- |
31
+ | `model.safetensors` | `BF16` | `517 MB` | Frozen text-only `google/embeddinggemma-2` weights (`language_model.*`; zero vision/audio weights) |
32
+ | `heads.safetensors` | `F32` | `58 MB` | Rank-16 LoRA (`layers [17, 20, 21, 22, 23]` + `embedding_projection`, bypassed during prompt embedding for `0.0` retrieval drift) + `ChoiceHead`, `NoulHead`, and `ScoreHead` |
33
+ | `heads.json` | JSON | `9.8 KB` | Head dimensions, query templates, LoRA metadata, and post-hoc `(T, b)` calibration tables |
34
+ | `calibration.json` | JSON | `2.6 KB` | Single-shot untouched `bar_locked` holdout seal |
35
+ | `tokenizer.json` | JSON | `31 MB` | Gemma fast BPE tokenizer |
36
+ | `config.json` | JSON | `5.0 KB` | Backbone text configuration |
37
+
38
+ ---
39
+
40
+ ## Untouched Holdout Calibration (`bar_locked`, `6,155` Rows)
41
+
42
+ Trained and calibrated on leak-free search and counterfactual multi-rollout trajectories (`0` root or query overlap with `bar_locked` and downstream evaluation benchmarks):
43
+
44
+ | Cell | Holdout Rows | Proper Loss (vs. Base Rate) | Accuracy / Conditional RPS | Quantile ECE (Gate `<= 0.045`) | Perturbation JS (Gate `<= 0.005`) |
45
+ | :--- | ---: | :--- | :--- | :--- | :--- |
46
+ | **`Choice:text`** | `2,024` | `0.1250 < 0.1599` | `49.2%` (`5`-way) | `0.0307` | `0.00079` |
47
+ | **`Noul:text`** | `2,040` | `0.1564 < 0.2500` | `77.5%` | `0.0315` | `0.00130` |
48
+ | **`Score:text`** | `2,091` | `0.1185 < 0.3340` | `RPS 0.1554 < 0.2004` | `0.0166` | `0.00130` |
49
+ | **Combined** | **`6,155`** | **All `3/3` Heads Admitted** | — | **`0.0262 <= 0.040`** | **`0.00113 <= 0.005`** |
50
+
51
+ ---
52
+
53
+ ## License & Attribution
54
+
55
+ - Base weights (`model.safetensors`, `tokenizer.json`, `config.json`) derive from [`google/embeddinggemma-2`](https://huggingface.co/google/embeddinggemma-2) under the **Apache-2.0** license.
56
+ - LoRA adapters, calibrated RLCD projection heads, and configuration metadata are released under **Apache-2.0**.
calibration.json ADDED
@@ -0,0 +1,104 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "version": "1.0",
3
+ "backbone": "google/embeddinggemma-2",
4
+ "admitted_heads": [
5
+ "Choice",
6
+ "Noul",
7
+ "Score"
8
+ ],
9
+ "all_text_cells_passed": true,
10
+ "overall_text_ece": 0.026192716122485877,
11
+ "overall_text_js": 0.001131477665816601,
12
+ "cells": {
13
+ "Choice:text": {
14
+ "rows": 2024,
15
+ "proper": 0.12495984882116318,
16
+ "baseRate": 0.15993022918701172,
17
+ "accuracy": 0.49160078167915344,
18
+ "ece": 0.03072668607585986,
19
+ "js": 0.0007881656638346612,
20
+ "admitted": true
21
+ },
22
+ "Noul:text": {
23
+ "rows": 2040,
24
+ "proper": 0.15640591084957123,
25
+ "baseRate": 0.25,
26
+ "accuracy": 0.7754902243614197,
27
+ "ece": 0.03153471350669861,
28
+ "js": 0.0012993740383535624,
29
+ "admitted": true
30
+ },
31
+ "Score:text": {
32
+ "rows": 2091,
33
+ "proper": 0.11847522854804993,
34
+ "baseRate": 0.3339491779521442,
35
+ "conditionalRps": 0.15541276335716248,
36
+ "conditionalRpsBase": 0.20038728415966034,
37
+ "ece": 0.016592319255234356,
38
+ "presenceEce": 2.93352910629703e-05,
39
+ "rankEce": 0.03315530321940574,
40
+ "js": 0.0012999878963455558,
41
+ "admitted": true
42
+ },
43
+ "Choice:image": {
44
+ "rows": 0,
45
+ "admitted": false,
46
+ "reason": "no_local_search_media_supervision_in_v1"
47
+ },
48
+ "Noul:image": {
49
+ "rows": 0,
50
+ "admitted": false,
51
+ "reason": "no_local_search_media_supervision_in_v1"
52
+ },
53
+ "Score:image": {
54
+ "rows": 0,
55
+ "admitted": false,
56
+ "reason": "no_local_search_media_supervision_in_v1"
57
+ },
58
+ "Extract:image": {
59
+ "rows": 0,
60
+ "admitted": false,
61
+ "reason": "no_local_search_media_supervision_in_v1"
62
+ },
63
+ "Choice:audio": {
64
+ "rows": 0,
65
+ "admitted": false,
66
+ "reason": "no_local_search_media_supervision_in_v1"
67
+ },
68
+ "Noul:audio": {
69
+ "rows": 0,
70
+ "admitted": false,
71
+ "reason": "no_local_search_media_supervision_in_v1"
72
+ },
73
+ "Score:audio": {
74
+ "rows": 0,
75
+ "admitted": false,
76
+ "reason": "no_local_search_media_supervision_in_v1"
77
+ },
78
+ "Extract:audio": {
79
+ "rows": 0,
80
+ "admitted": false,
81
+ "reason": "no_local_search_media_supervision_in_v1"
82
+ },
83
+ "Choice:video": {
84
+ "rows": 0,
85
+ "admitted": false,
86
+ "reason": "no_local_search_media_supervision_in_v1"
87
+ },
88
+ "Noul:video": {
89
+ "rows": 0,
90
+ "admitted": false,
91
+ "reason": "no_local_search_media_supervision_in_v1"
92
+ },
93
+ "Score:video": {
94
+ "rows": 0,
95
+ "admitted": false,
96
+ "reason": "no_local_search_media_supervision_in_v1"
97
+ },
98
+ "Extract:video": {
99
+ "rows": 0,
100
+ "admitted": false,
101
+ "reason": "no_local_search_media_supervision_in_v1"
102
+ }
103
+ }
104
+ }
config.json ADDED
@@ -0,0 +1,208 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "transformers_version": "5.23.0",
3
+ "architectures": [
4
+ "EmbeddingGemma2Model"
5
+ ],
6
+ "output_hidden_states": false,
7
+ "return_dict": true,
8
+ "dtype": "float32",
9
+ "chunk_size_feed_forward": 0,
10
+ "is_encoder_decoder": false,
11
+ "id2label": {
12
+ "0": "LABEL_0",
13
+ "1": "LABEL_1"
14
+ },
15
+ "label2id": {
16
+ "LABEL_0": 0,
17
+ "LABEL_1": 1
18
+ },
19
+ "problem_type": null,
20
+ "text_config": {
21
+ "architectures": null,
22
+ "output_hidden_states": false,
23
+ "return_dict": true,
24
+ "dtype": "float32",
25
+ "chunk_size_feed_forward": 0,
26
+ "is_encoder_decoder": false,
27
+ "id2label": {
28
+ "0": "LABEL_0",
29
+ "1": "LABEL_1"
30
+ },
31
+ "label2id": {
32
+ "LABEL_0": 0,
33
+ "LABEL_1": 1
34
+ },
35
+ "problem_type": null,
36
+ "vocab_size": 262144,
37
+ "hidden_size": 512,
38
+ "intermediate_size": 2048,
39
+ "num_hidden_layers": 24,
40
+ "num_attention_heads": 4,
41
+ "num_key_value_heads": 2,
42
+ "head_dim": 256,
43
+ "hidden_activation": "gelu_pytorch_tanh",
44
+ "max_position_embeddings": 262144,
45
+ "initializer_range": 0.02,
46
+ "rms_norm_eps": 1e-06,
47
+ "pad_token_id": 0,
48
+ "eos_token_id": 1,
49
+ "bos_token_id": 2,
50
+ "rope_parameters": {
51
+ "full_attention": {
52
+ "rope_theta": 1000000.0,
53
+ "rope_type": "default"
54
+ },
55
+ "sliding_attention": {
56
+ "rope_theta": 10000.0,
57
+ "rope_type": "default"
58
+ }
59
+ },
60
+ "attention_bias": false,
61
+ "attention_dropout": 0.0,
62
+ "sliding_window": 512,
63
+ "layer_types": [
64
+ "sliding_attention",
65
+ "sliding_attention",
66
+ "sliding_attention",
67
+ "sliding_attention",
68
+ "sliding_attention",
69
+ "full_attention",
70
+ "sliding_attention",
71
+ "sliding_attention",
72
+ "sliding_attention",
73
+ "sliding_attention",
74
+ "sliding_attention",
75
+ "full_attention",
76
+ "sliding_attention",
77
+ "sliding_attention",
78
+ "sliding_attention",
79
+ "sliding_attention",
80
+ "sliding_attention",
81
+ "full_attention",
82
+ "sliding_attention",
83
+ "sliding_attention",
84
+ "sliding_attention",
85
+ "sliding_attention",
86
+ "sliding_attention",
87
+ "full_attention"
88
+ ],
89
+ "hidden_size_per_layer_input": 512,
90
+ "embedding_dim": 768,
91
+ "_name_or_path": "",
92
+ "model_type": "embedding_gemma2_text",
93
+ "output_attentions": false,
94
+ "per_layer_config": {
95
+ "05": {
96
+ "head_dim": 512,
97
+ "num_key_value_heads": 1
98
+ },
99
+ "11": {
100
+ "head_dim": 512,
101
+ "num_key_value_heads": 1
102
+ },
103
+ "17": {
104
+ "head_dim": 512,
105
+ "num_key_value_heads": 1
106
+ },
107
+ "23": {
108
+ "head_dim": 512,
109
+ "num_key_value_heads": 1
110
+ }
111
+ }
112
+ },
113
+ "vision_config": {
114
+ "architectures": null,
115
+ "output_hidden_states": false,
116
+ "return_dict": true,
117
+ "dtype": "float32",
118
+ "chunk_size_feed_forward": 0,
119
+ "is_encoder_decoder": false,
120
+ "id2label": {
121
+ "0": "LABEL_0",
122
+ "1": "LABEL_1"
123
+ },
124
+ "label2id": {
125
+ "LABEL_0": 0,
126
+ "LABEL_1": 1
127
+ },
128
+ "problem_type": null,
129
+ "hidden_size": 768,
130
+ "intermediate_size": 3072,
131
+ "num_hidden_layers": 16,
132
+ "num_attention_heads": 12,
133
+ "num_key_value_heads": 12,
134
+ "head_dim": 64,
135
+ "hidden_activation": "gelu_pytorch_tanh",
136
+ "rms_norm_eps": 1e-06,
137
+ "max_position_embeddings": 131072,
138
+ "attention_bias": false,
139
+ "attention_dropout": 0.0,
140
+ "rope_parameters": {
141
+ "rope_theta": 100.0,
142
+ "rope_type": "axial"
143
+ },
144
+ "pooling_kernel_size": 3,
145
+ "patch_size": 16,
146
+ "position_embedding_size": 10240,
147
+ "use_clipped_linears": false,
148
+ "standardize": false,
149
+ "initializer_range": 0.02,
150
+ "_name_or_path": "",
151
+ "default_output_length": 280,
152
+ "global_head_dim": 64,
153
+ "model_type": "gemma4_vision",
154
+ "output_attentions": false
155
+ },
156
+ "audio_config": {
157
+ "architectures": null,
158
+ "output_hidden_states": false,
159
+ "return_dict": true,
160
+ "dtype": "float32",
161
+ "chunk_size_feed_forward": 0,
162
+ "is_encoder_decoder": false,
163
+ "id2label": {
164
+ "0": "LABEL_0",
165
+ "1": "LABEL_1"
166
+ },
167
+ "label2id": {
168
+ "LABEL_0": 0,
169
+ "LABEL_1": 1
170
+ },
171
+ "problem_type": null,
172
+ "hidden_size": 1024,
173
+ "num_hidden_layers": 12,
174
+ "num_attention_heads": 8,
175
+ "hidden_act": "silu",
176
+ "subsampling_conv_channels": [
177
+ 128,
178
+ 32
179
+ ],
180
+ "conv_kernel_size": 5,
181
+ "residual_weight": 0.5,
182
+ "attention_chunk_size": 12,
183
+ "attention_context_left": 13,
184
+ "attention_context_right": 0,
185
+ "attention_logit_cap": 50.0,
186
+ "attention_invalid_logits_value": -1000000000.0,
187
+ "use_clipped_linears": true,
188
+ "rms_norm_eps": 1e-06,
189
+ "gradient_clipping": 10000000000.0,
190
+ "output_proj_dims": 1536,
191
+ "initializer_range": 0.02,
192
+ "_name_or_path": "",
193
+ "model_type": "gemma4_audio",
194
+ "output_attentions": false
195
+ },
196
+ "boi_token_id": 255999,
197
+ "eoi_token_id": 258882,
198
+ "image_token_id": 258880,
199
+ "video_token_id": 258884,
200
+ "boa_token_id": 256000,
201
+ "eoa_token_index": 258883,
202
+ "audio_token_id": 258881,
203
+ "initializer_range": 0.02,
204
+ "_name_or_path": "google/embeddinggemma-2",
205
+ "model_type": "embedding_gemma2",
206
+ "vision_soft_tokens_per_image": 280,
207
+ "output_attentions": false
208
+ }
heads.json ADDED
@@ -0,0 +1,326 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "d": 768,
3
+ "num_score_metrics": 3,
4
+ "num_ranks": 5,
5
+ "include_extract": false,
6
+ "active_heads": [
7
+ "choice",
8
+ "noul",
9
+ "score"
10
+ ],
11
+ "best_epoch": 6,
12
+ "query_texts": {
13
+ "choice_next_step": "task: classification | query: Given the research intent and current parallel search returns, which next action should the search runtime take?",
14
+ "choice_best_query_for_intent": "task: classification | query: Which candidate search query is most specific, well-formed, and likely to surface authoritative evidence for the research intent?",
15
+ "choice_best_search_hit": "task: classification | query: Which candidate search hit provides the strongest authoritative evidence to answer the task?",
16
+ "noul_query_matches_intent": "task: fact checking | query: This candidate search sub-query is well-formed, specific to the research intent, and likely to surface useful evidence.",
17
+ "noul_query_returns_results": "task: fact checking | query: This search query returns one or more web results from the Search API.",
18
+ "noul_query_is_fact_shaped": "task: fact checking | query: This search query asks for a structured quantitative, financial, statistical, or live metric fact covered by licensed knowledge providers.",
19
+ "noul_results_relevant": "task: fact checking | query: The returned search results contain relevant information that addresses the search query.",
20
+ "noul_hit_worth_reading": "task: fact checking | query: This candidate search hit is worth fetching in full to answer the user task.",
21
+ "score_return_relevance": "task: classification | query: Rate how directly and authoritatively this search return (web highlight or knowledge fact) helps answer the research intent from 1 to 5.",
22
+ "score_evidence_coverage": "task: classification | query: Rate how completely this fused set of parallel search highlights and knowledge facts covers all required parts of the research intent from 1 to 5.",
23
+ "score_extract_relevance": "task: classification | query: Rate how directly and completely this extracted webpage or PDF passage satisfies the document intent from 1 to 5."
24
+ },
25
+ "score_metrics": [
26
+ "score_return_relevance",
27
+ "score_evidence_coverage",
28
+ "score_extract_relevance"
29
+ ],
30
+ "calibration": {
31
+ "choice": {
32
+ "choice_best_query_for_intent": {
33
+ "T": 1.3695668877685563,
34
+ "b_abstain": -0.058805350214242935
35
+ },
36
+ "choice_best_search_hit": {
37
+ "T": 1.817622866872739,
38
+ "b_abstain": -0.32487934827804565
39
+ },
40
+ "choice_next_step": {
41
+ "T": 0.5006811022758484,
42
+ "b_abstain": -0.0876120999455452
43
+ }
44
+ },
45
+ "noul": {
46
+ "noul_query_returns_results": {
47
+ "T": 2.9460168033182867,
48
+ "b": -1.4611535458463436
49
+ },
50
+ "noul_query_matches_intent": {
51
+ "T": 1.3391732586857756,
52
+ "b": -0.8235207292718201
53
+ },
54
+ "noul_hit_worth_reading": {
55
+ "T": 0.8524871718440892,
56
+ "b": 0.28758746344588815
57
+ }
58
+ },
59
+ "score": {
60
+ "score_return_relevance": {
61
+ "T_pres": 0.4000000059604645,
62
+ "b_pres": 0.6136154532432556,
63
+ "T_rank": 1.917371153831482,
64
+ "b_rank": [
65
+ 0.09833735227584839,
66
+ 0.039705585688352585,
67
+ -0.02585119754076004,
68
+ -0.002763827098533511,
69
+ -0.10942791402339935
70
+ ]
71
+ },
72
+ "score_evidence_coverage": {
73
+ "T_pres": 0.4000000059604645,
74
+ "b_pres": 0.6011375784873962,
75
+ "T_rank": 1.4472062587738037,
76
+ "b_rank": [
77
+ -0.18481789529323578,
78
+ -0.08214931190013885,
79
+ 0.020597977563738823,
80
+ 0.08634351938962936,
81
+ 0.1600256860256195
82
+ ]
83
+ },
84
+ "score_extract_relevance": {
85
+ "T_pres": 0.4216231107711792,
86
+ "b_pres": 0.5255321860313416,
87
+ "T_rank": 2.2792673110961914,
88
+ "b_rank": [
89
+ 0.031077247112989426,
90
+ -0.023274000734090805,
91
+ -0.036772601306438446,
92
+ 0.030096551403403282,
93
+ -0.0011271957773715258
94
+ ]
95
+ }
96
+ },
97
+ "extract": {}
98
+ },
99
+ "lora": {
100
+ "r": 16,
101
+ "alpha": 32.0,
102
+ "layer_indices": [
103
+ 17,
104
+ 20,
105
+ 21,
106
+ 22,
107
+ 23
108
+ ],
109
+ "include_projection": true
110
+ },
111
+ "backbone_model_id": "google/embeddinggemma-2",
112
+ "backbone_config": {
113
+ "transformers_version": "5.23.0",
114
+ "architectures": [
115
+ "EmbeddingGemma2Model"
116
+ ],
117
+ "output_hidden_states": false,
118
+ "return_dict": true,
119
+ "dtype": "float32",
120
+ "chunk_size_feed_forward": 0,
121
+ "is_encoder_decoder": false,
122
+ "id2label": {
123
+ "0": "LABEL_0",
124
+ "1": "LABEL_1"
125
+ },
126
+ "label2id": {
127
+ "LABEL_0": 0,
128
+ "LABEL_1": 1
129
+ },
130
+ "problem_type": null,
131
+ "text_config": {
132
+ "architectures": null,
133
+ "output_hidden_states": false,
134
+ "return_dict": true,
135
+ "dtype": "float32",
136
+ "chunk_size_feed_forward": 0,
137
+ "is_encoder_decoder": false,
138
+ "id2label": {
139
+ "0": "LABEL_0",
140
+ "1": "LABEL_1"
141
+ },
142
+ "label2id": {
143
+ "LABEL_0": 0,
144
+ "LABEL_1": 1
145
+ },
146
+ "problem_type": null,
147
+ "vocab_size": 262144,
148
+ "hidden_size": 512,
149
+ "intermediate_size": 2048,
150
+ "num_hidden_layers": 24,
151
+ "num_attention_heads": 4,
152
+ "num_key_value_heads": 2,
153
+ "head_dim": 256,
154
+ "hidden_activation": "gelu_pytorch_tanh",
155
+ "max_position_embeddings": 262144,
156
+ "initializer_range": 0.02,
157
+ "rms_norm_eps": 1e-06,
158
+ "pad_token_id": 0,
159
+ "eos_token_id": 1,
160
+ "bos_token_id": 2,
161
+ "rope_parameters": {
162
+ "full_attention": {
163
+ "rope_theta": 1000000.0,
164
+ "rope_type": "default"
165
+ },
166
+ "sliding_attention": {
167
+ "rope_theta": 10000.0,
168
+ "rope_type": "default"
169
+ }
170
+ },
171
+ "attention_bias": false,
172
+ "attention_dropout": 0.0,
173
+ "sliding_window": 512,
174
+ "layer_types": [
175
+ "sliding_attention",
176
+ "sliding_attention",
177
+ "sliding_attention",
178
+ "sliding_attention",
179
+ "sliding_attention",
180
+ "full_attention",
181
+ "sliding_attention",
182
+ "sliding_attention",
183
+ "sliding_attention",
184
+ "sliding_attention",
185
+ "sliding_attention",
186
+ "full_attention",
187
+ "sliding_attention",
188
+ "sliding_attention",
189
+ "sliding_attention",
190
+ "sliding_attention",
191
+ "sliding_attention",
192
+ "full_attention",
193
+ "sliding_attention",
194
+ "sliding_attention",
195
+ "sliding_attention",
196
+ "sliding_attention",
197
+ "sliding_attention",
198
+ "full_attention"
199
+ ],
200
+ "hidden_size_per_layer_input": 512,
201
+ "embedding_dim": 768,
202
+ "_name_or_path": "",
203
+ "model_type": "embedding_gemma2_text",
204
+ "output_attentions": false,
205
+ "per_layer_config": {
206
+ "05": {
207
+ "head_dim": 512,
208
+ "num_key_value_heads": 1
209
+ },
210
+ "11": {
211
+ "head_dim": 512,
212
+ "num_key_value_heads": 1
213
+ },
214
+ "17": {
215
+ "head_dim": 512,
216
+ "num_key_value_heads": 1
217
+ },
218
+ "23": {
219
+ "head_dim": 512,
220
+ "num_key_value_heads": 1
221
+ }
222
+ }
223
+ },
224
+ "vision_config": {
225
+ "architectures": null,
226
+ "output_hidden_states": false,
227
+ "return_dict": true,
228
+ "dtype": "float32",
229
+ "chunk_size_feed_forward": 0,
230
+ "is_encoder_decoder": false,
231
+ "id2label": {
232
+ "0": "LABEL_0",
233
+ "1": "LABEL_1"
234
+ },
235
+ "label2id": {
236
+ "LABEL_0": 0,
237
+ "LABEL_1": 1
238
+ },
239
+ "problem_type": null,
240
+ "hidden_size": 768,
241
+ "intermediate_size": 3072,
242
+ "num_hidden_layers": 16,
243
+ "num_attention_heads": 12,
244
+ "num_key_value_heads": 12,
245
+ "head_dim": 64,
246
+ "hidden_activation": "gelu_pytorch_tanh",
247
+ "rms_norm_eps": 1e-06,
248
+ "max_position_embeddings": 131072,
249
+ "attention_bias": false,
250
+ "attention_dropout": 0.0,
251
+ "rope_parameters": {
252
+ "rope_theta": 100.0,
253
+ "rope_type": "axial"
254
+ },
255
+ "pooling_kernel_size": 3,
256
+ "patch_size": 16,
257
+ "position_embedding_size": 10240,
258
+ "use_clipped_linears": false,
259
+ "standardize": false,
260
+ "initializer_range": 0.02,
261
+ "_name_or_path": "",
262
+ "default_output_length": 280,
263
+ "global_head_dim": 64,
264
+ "model_type": "gemma4_vision",
265
+ "output_attentions": false
266
+ },
267
+ "audio_config": {
268
+ "architectures": null,
269
+ "output_hidden_states": false,
270
+ "return_dict": true,
271
+ "dtype": "float32",
272
+ "chunk_size_feed_forward": 0,
273
+ "is_encoder_decoder": false,
274
+ "id2label": {
275
+ "0": "LABEL_0",
276
+ "1": "LABEL_1"
277
+ },
278
+ "label2id": {
279
+ "LABEL_0": 0,
280
+ "LABEL_1": 1
281
+ },
282
+ "problem_type": null,
283
+ "hidden_size": 1024,
284
+ "num_hidden_layers": 12,
285
+ "num_attention_heads": 8,
286
+ "hidden_act": "silu",
287
+ "subsampling_conv_channels": [
288
+ 128,
289
+ 32
290
+ ],
291
+ "conv_kernel_size": 5,
292
+ "residual_weight": 0.5,
293
+ "attention_chunk_size": 12,
294
+ "attention_context_left": 13,
295
+ "attention_context_right": 0,
296
+ "attention_logit_cap": 50.0,
297
+ "attention_invalid_logits_value": -1000000000.0,
298
+ "use_clipped_linears": true,
299
+ "rms_norm_eps": 1e-06,
300
+ "gradient_clipping": 10000000000.0,
301
+ "output_proj_dims": 1536,
302
+ "initializer_range": 0.02,
303
+ "_name_or_path": "",
304
+ "model_type": "gemma4_audio",
305
+ "output_attentions": false
306
+ },
307
+ "boi_token_id": 255999,
308
+ "eoi_token_id": 258882,
309
+ "image_token_id": 258880,
310
+ "video_token_id": 258884,
311
+ "boa_token_id": 256000,
312
+ "eoa_token_index": 258883,
313
+ "audio_token_id": 258881,
314
+ "initializer_range": 0.02,
315
+ "_name_or_path": "google/embeddinggemma-2",
316
+ "model_type": "embedding_gemma2",
317
+ "vision_soft_tokens_per_image": 280,
318
+ "output_attentions": false
319
+ },
320
+ "special_tokens": {
321
+ "bos_token": "<bos>",
322
+ "eos_token": "<eos>",
323
+ "pad_token": "<pad>",
324
+ "unk_token": "<unk>"
325
+ }
326
+ }
heads.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:03f1f4a68eb9eb74d3dfc9e77437e8f318ad94f67401bde36f3cc92358acc47e
3
+ size 61226932
model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:8206ed235a4c5441afedacdeea628a372519895cf3521ce7b5702a83f909afb9
3
+ size 542056416
tokenizer.json ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4d777ef5bdc1aa36227abdfb77c3e49e7b9c892d16e1b6bda41c393504828be4
3
+ size 32170510