rishanthrajendhran commited on
Commit
cc9ecee
·
verified ·
1 Parent(s): bd12292

Release: merged weights, adapter/ (Tinker LoRA), load_adapter.py, thresholds, card, LICENSE

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
LICENSE ADDED
@@ -0,0 +1,51 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
2
+
3
+ OpenMDW License Agreement, version 1.1 (OpenMDW-1.1)
4
+
5
+ By exercising rights granted to you under this agreement, you accept and agree
6
+ to its terms.
7
+
8
+ As used in this agreement, "Model Materials" means the materials provided to
9
+ you under this agreement, consisting of: (1) one or more machine learning
10
+ models (including architecture and parameters); and (2) all related artifacts
11
+ (including associated data, documentation and software) that are provided to
12
+ you hereunder.
13
+
14
+ Subject to your compliance with this agreement, permission is hereby granted,
15
+ free of charge, to deal in the Model Materials without restriction, including
16
+ under all copyright, patent, database, and trade secret rights included or
17
+ embodied therein.
18
+
19
+ If you distribute any portion of the Model Materials, you shall retain in your
20
+ distribution (1) a copy of this agreement, and (2) all copyright notices and
21
+ other notices of origin included in the Model Materials that are applicable to
22
+ your distribution.
23
+
24
+ If you file, maintain, or voluntarily participate in a lawsuit against any
25
+ person or entity asserting that the Model Materials directly or indirectly
26
+ infringe any patent or copyright, then all rights and grants made to you
27
+ hereunder are terminated, unless that lawsuit was in response to a
28
+ corresponding lawsuit first brought against you.
29
+
30
+ This agreement does not impose any restrictions or obligations with respect to
31
+ any use, modification, or sharing of any outputs generated by using the Model
32
+ Materials.
33
+
34
+ THE MODEL MATERIALS ARE PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
35
+ OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
36
+ FITNESS FOR A PARTICULAR PURPOSE, TITLE, NONINFRINGEMENT, ACCURACY, OR THE
37
+ ABSENCE OF LATENT OR OTHER DEFECTS OR ERRORS, WHETHER OR NOT DISCOVERABLE, ALL
38
+ TO THE GREATEST EXTENT PERMISSIBLE UNDER APPLICABLE LAW.
39
+
40
+ YOU ARE SOLELY RESPONSIBLE FOR (1) CLEARING RIGHTS OF OTHER PERSONS THAT MAY
41
+ APPLY TO THE MODEL MATERIALS OR ANY USE THEREOF, INCLUDING WITHOUT LIMITATION
42
+ ANY PERSON'S COPYRIGHTS OR OTHER RIGHTS INCLUDED OR EMBODIED IN THE MODEL
43
+ MATERIALS; (2) OBTAINING ANY NECESSARY CONSENTS, PERMISSIONS OR OTHER RIGHTS
44
+ REQUIRED FOR ANY USE OF THE MODEL MATERIALS; OR (3) PERFORMING ANY DUE
45
+ DILIGENCE OR UNDERTAKING ANY OTHER INVESTIGATIONS INTO THE MODEL MATERIALS OR
46
+ ANYTHING INCORPORATED OR EMBODIED THEREIN.
47
+
48
+ IN NO EVENT SHALL THE PROVIDERS OF THE MODEL MATERIALS BE LIABLE FOR ANY CLAIM,
49
+ DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
50
+ OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE MODEL MATERIALS, THE
51
+ USE THEREOF OR OTHER DEALINGS THEREIN.
README.md CHANGED
@@ -1,24 +1,194 @@
1
  ---
2
  license: other
3
  license_name: openmdw-1.1
 
4
  base_model: nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16
5
- library_name: peft
6
- tags: [ai-detection, lora]
 
 
 
 
 
 
7
  extra_gated_prompt: "Access is granted individually. Please say who you are and what you intend to use the weights for."
8
  ---
9
 
10
- # ideadet-nemotron30b-1m-document
11
 
12
- Idea-level AI-text detector: a LoRA adapter over
13
- `nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16`. It answers **whose ideas a document
14
- contains**, not who typed the sentences.
15
 
16
- - **Training corpus:** 1m
17
- - **Reads at inference:** the raw document text
18
 
19
- Scored by comparing the next-token log-probabilities of the tokens `human` and `ai`, two-way
20
- normalised, giving P(human). The detector fires when that falls below a calibrated threshold.
21
- Thresholds are quantiles over held-out human documents and are **not** included here: a cut
22
- from one model or input form is meaningless against another's scores.
23
 
24
- Fuller documentation and evaluation results to follow.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
2
  license: other
3
  license_name: openmdw-1.1
4
+ license_link: LICENSE
5
  base_model: nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16
6
+ library_name: transformers
7
+ language:
8
+ - en
9
+ tags:
10
+ - ai-text-detection
11
+ - prose-provenance
12
+ datasets:
13
+ - rishanthrajendhran/IdeaLens-1M
14
  extra_gated_prompt: "Access is granted individually. Please say who you are and what you intend to use the weights for."
15
  ---
16
 
17
+ # ProseLens
18
 
19
+ ProseLens detects **who wrote the words** of a document. It is IdeaLens's counterpart in the paper: the same
20
+ backbone, training documents and labels, but it reads the raw document text instead of an outline, so it learns
21
+ word-level provenance. It returns P(human), the probability that the document was written by a person.
22
 
23
+ ProseLens is `nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16` fine-tuned with LoRA (rank 64) on 1M English web
24
+ documents ([IdeaLens-1M](https://huggingface.co/datasets/rishanthrajendhran/IdeaLens-1M)).
25
 
26
+ ## Results
 
 
 
27
 
28
+ From the paper: ProseLens is accurate when a document's ideas and words come from the same source (99.1%), but
29
+ when they come from different sources it tracks the words and reaches 25.4% on idea provenance. IdeaLens, trained
30
+ identically on outlines, reaches 95.3% and 81.3%; Pangram 4 reaches 98.5% and 25.9%.
31
+
32
+ ## Usage
33
+
34
+ Pass the document text as is.
35
+
36
+ ### Load the merged model (66 GB download)
37
+
38
+ ```python
39
+ import torch
40
+ from transformers import AutoModelForCausalLM, AutoTokenizer
41
+
42
+ tok = AutoTokenizer.from_pretrained("rishanthrajendhran/ProseLens")
43
+ model = AutoModelForCausalLM.from_pretrained("rishanthrajendhran/ProseLens", dtype=torch.bfloat16, device_map="auto").eval()
44
+ ```
45
+
46
+ The weights take 59 GiB of GPU memory, and each input adds more; see *Hardware requirements*.
47
+
48
+ ### Or apply the adapter to the base model (3 GB download)
49
+
50
+ `adapter/` holds the LoRA adapter as trained, in the layout of the Tinker training service. If you already have the
51
+ base model, `load_adapter.py` merges the adapter into it in memory. The resulting weights are bit-identical to the
52
+ merged model's:
53
+
54
+ ```python
55
+ import importlib.util
56
+ from huggingface_hub import hf_hub_download
57
+
58
+ path = hf_hub_download("rishanthrajendhran/ProseLens", "load_adapter.py")
59
+ spec = importlib.util.spec_from_file_location("load_adapter", path)
60
+ la = importlib.util.module_from_spec(spec); spec.loader.exec_module(la)
61
+ model, tok = la.load_model() # base model + adapter/, then la.p_human(model, tok, text)
62
+ ```
63
+
64
+ Do not load `adapter/` with `peft.PeftModel`. In transformers, Nemotron fuses the Mamba gate and x projections into
65
+ one `in_proj` and stores each layer's 128 routed experts as a single 3D tensor, so PEFT has nowhere to attach most of
66
+ the adapter and skips it without a warning; the model then scores close to the base model. `tinker-cookbook`'s
67
+ `weights.build_hf_model` can also merge the adapter into full weights.
68
+
69
+ ### Score a document
70
+
71
+ ProseLens compares the next-token probabilities of `human` and `ai`:
72
+
73
+ ```python
74
+ SYSTEM = "Given a document, answer with one word: human if the document was human-written, ai if it was AI-generated."
75
+ SUFFIX = "<|im_end|>\n<|im_start|>assistant\n<think>\n\n</think>\n\n"
76
+ HUMAN, AI = 50755, 2464 # token ids of "human" and "ai"
77
+
78
+ @torch.no_grad()
79
+ def p_human(text):
80
+ ids = tok.encode(f"<|im_start|>system\n{SYSTEM}<|im_end|>\n<|im_start|>user\n{text}{SUFFIX}",
81
+ add_special_tokens=False)
82
+ logits = model(torch.tensor([ids], device=model.device)).logits[0, -1].float()
83
+ return torch.softmax(logits[[HUMAN, AI]], -1)[0].item()
84
+
85
+ text = open("document.txt").read()
86
+ print(p_human(text))
87
+ ```
88
+
89
+ Build the prompt string exactly as above rather than through the chat template.
90
+
91
+ ### Thresholds
92
+
93
+ ProseLens flags a document as AI-written when P(human) is at or below a cut. Each cut is set so
94
+ that a given share of human documents is flagged (the false-positive rate, FPR), measured on the 80,000 human
95
+ documents in IdeaLens-1M's `calibration` split (10,000 per format). The paper's operating point is the global
96
+ cut at 1% FPR.
97
+
98
+ | FPR | 0.1% | 0.5% | 1% | 2% | 5% |
99
+ |---|---:|---:|---:|---:|---:|
100
+ | Global cut | 0.07082 | 0.30469 | 0.60539 | 0.92502 | 0.99889 |
101
+
102
+ Per-format cuts give each format its own operating point. They need the document's format, which the paper
103
+ assigns with WebOrganizer's FormatClassifier (FormatClassifier-NoURL when there is no URL). Each is the
104
+ format's own quantile, shrunk toward the global cut with weight n / (n + 2500); at 0.1% FPR 10,000 documents
105
+ per format are too few, so there is no per-format cut. A document outside these eight formats has no
106
+ per-format cut; do not fall back to the global cut for it.
107
+
108
+ | Format | 0.5% | 1% | 2% | 5% |
109
+ |---|---:|---:|---:|---:|
110
+ | Nonfiction Writing | 0.14387 | 0.27050 | 0.49499 | 0.93935 |
111
+ | Knowledge Article | 0.24236 | 0.38050 | 0.64044 | 0.95295 |
112
+ | Personal Blog | 0.68275 | 0.87993 | 0.98163 | 0.99966 |
113
+ | News Article | 0.29554 | 0.55535 | 0.90380 | 0.99788 |
114
+ | Academic Writing | 0.77788 | 0.89967 | 0.98279 | 0.99969 |
115
+ | User Reviews | 0.62135 | 0.85161 | 0.98035 | 0.99964 |
116
+ | Personal About Page | 0.41699 | 0.78714 | 0.97730 | 0.99966 |
117
+ | Creative Writing | 0.84923 | 0.91749 | 0.98411 | 0.99958 |
118
+
119
+ `thresholds.json` holds every cut at full precision.
120
+
121
+ These rates hold for English web documents like the training data. For another domain, fit the cut on
122
+ human documents from that domain.
123
+
124
+ ProseLens scores 97% of human calibration documents above 0.99, so its cuts at higher FPRs sit close
125
+ to 1 and small shifts in score move the realised FPR a long way.
126
+
127
+ ## Hardware requirements
128
+
129
+ Measured with transformers 5.15 in bf16 on NVIDIA H100 80GB GPUs (our other runs used A100 80GB), with transformers'
130
+ PyTorch implementation of the Mamba layers (no fused Mamba kernels installed). We have not tried CPU-only inference.
131
+
132
+ | | Merged model | Adapter route (`load_adapter.py`) |
133
+ |---|---|---|
134
+ | Download | 65.8 GB | 65.8 GB base model + 3.1 GB adapter |
135
+ | Peak CPU RAM while loading | 60 GiB | 60 GiB |
136
+ | GPU memory once loaded | 58.8 GiB | 58.8 GiB (66 GiB during the ~10 s it takes to apply the adapter) |
137
+
138
+ GPU memory then grows with the length of the input, by about 4.2 MiB per token at typical lengths, scoring one
139
+ input at a time:
140
+
141
+ | Input tokens | 500 | 1,000 | 2,000 | 4,000 | 8,000 |
142
+ |---|---:|---:|---:|---:|---:|
143
+ | Peak GPU memory, one 80 GB GPU | 61.0 GiB | 63.1 GiB | 67.3 GiB | 75.7 GiB | does not fit |
144
+ | Peak memory per GPU, two 80 GB GPUs (`device_map="auto"`) | | | | 47.4 GiB | 63.7 GiB |
145
+ | Seconds per input, H100 | 0.18 | 0.34 | 0.66 | 1.32 | 2.70 |
146
+
147
+ Inputs of 6,000 tokens do not fit on one 80 GB GPU and 12,000 do not fit on two; lowering the Mamba chunk size from
148
+ 128 to 64 did not change either limit.
149
+
150
+ ProseLens reads whole documents, so their length decides the hardware. One 80 GB GPU handles documents up to about
151
+ 4,000 tokens (about 3,000 words); 85% of IdeaLens-1M's test documents are that short. Two 80 GB GPUs handle up to
152
+ 8,000 tokens (about 6,000 words), which covers all but about 1 in 1,000 test documents. We have not measured longer
153
+ inputs or more GPUs.
154
+
155
+ ## Intended use and limitations
156
+
157
+ - ProseLens estimates who wrote a document's words. It should not be the sole basis for decisions about a person's work.
158
+ - It was trained on English web documents of at least 500 words in eight long-form formats (Nonfiction Writing, Knowledge Article, Personal Blog, News Article, Academic Writing, User Reviews, Personal About Page, Creative Writing).
159
+ - Its training labels come from the Pangram prose detector, applied to whole documents.
160
+
161
+ ## Related models
162
+
163
+ | Model | Backbone | Reads |
164
+ |---|---|---|
165
+ | [IdeaLens](https://huggingface.co/rishanthrajendhran/IdeaLens) | Nemotron-3.5-Lightning-30B-A3B, LoRA | outline |
166
+ | [ProseLens](https://huggingface.co/rishanthrajendhran/ProseLens) (this model) | Nemotron-3.5-Lightning-30B-A3B, LoRA | document text |
167
+ | [IdeaLens-NoParaphrase](https://huggingface.co/rishanthrajendhran/IdeaLens-NoParaphrase) | Nemotron-3.5-Lightning-30B-A3B, LoRA | outline, trained without paraphrasing |
168
+ | [IdeaLens-Qwen3.5-9B](https://huggingface.co/rishanthrajendhran/IdeaLens-Qwen3.5-9B) | Qwen3.5-9B, classification head | outline |
169
+ | [IdeaLens-ModernBERT-L](https://huggingface.co/rishanthrajendhran/IdeaLens-ModernBERT-L) | ModernBERT-large | outline |
170
+ | [ProseLens-ModernBERT-L](https://huggingface.co/rishanthrajendhran/ProseLens-ModernBERT-L) | ModernBERT-large | document text |
171
+ | [IdeaLens-LogisticClassifier](https://huggingface.co/rishanthrajendhran/IdeaLens-LogisticClassifier) | logistic regression over text-embedding-3-large | outline |
172
+ | [IdeaLens-ModernBERT-L-NoParaphrase](https://huggingface.co/rishanthrajendhran/IdeaLens-ModernBERT-L-NoParaphrase) | ModernBERT-large | outline, trained without paraphrasing |
173
+ | [IdeaLens-ModernBERT-L-RolesOnly](https://huggingface.co/rishanthrajendhran/IdeaLens-ModernBERT-L-RolesOnly) | ModernBERT-large | role labels only |
174
+ | [IdeaLens-Qwen3.5-9B-PerItem](https://huggingface.co/rishanthrajendhran/IdeaLens-Qwen3.5-9B-PerItem) | Qwen3.5-9B, classification head | single outline items, pooled |
175
+ | [IdeaLens-ModernBERT-L-PerItem](https://huggingface.co/rishanthrajendhran/IdeaLens-ModernBERT-L-PerItem) | ModernBERT-large | single outline items, pooled |
176
+ | [IdeaLens-LogisticClassifier-PerItem](https://huggingface.co/rishanthrajendhran/IdeaLens-LogisticClassifier-PerItem) | logistic regression over text-embedding-3-large | single outline items, pooled |
177
+
178
+ Training data: [IdeaLens-1M](https://huggingface.co/datasets/rishanthrajendhran/IdeaLens-1M).
179
+
180
+ ## License
181
+
182
+ OpenMDW-1.1, the license of the base model (see `LICENSE`).
183
+
184
+ ## Citation
185
+
186
+ ```bibtex
187
+ @article{idealens2026,
188
+ title = {IdeaLens: Detecting AI Ideas in Long-form Writing},
189
+ author = {Anonymous},
190
+ journal = {arXiv preprint arXiv:TBD},
191
+ year = {2026},
192
+ url = {https://arxiv.org/abs/TBD}
193
+ }
194
+ ```
adapter_config.json → adapter/adapter_config.json RENAMED
File without changes
adapter_model.safetensors → adapter/adapter_model.safetensors RENAMED
File without changes
chat_template.jinja ADDED
@@ -0,0 +1,190 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {% macro render_extra_keys(json_dict, handled_keys) %}
2
+ {%- if json_dict is mapping %}
3
+ {%- for json_key in json_dict if json_key not in handled_keys %}
4
+ {%- if json_dict[json_key] is mapping or (json_dict[json_key] is sequence and json_dict[json_key] is not string) %}
5
+ {{- '\n<' ~ json_key ~ '>' ~ (json_dict[json_key] | tojson | safe) ~ '</' ~ json_key ~ '>' }}
6
+ {%- else %}
7
+ {{-'\n<' ~ json_key ~ '>' ~ (json_dict[json_key] | string) ~ '</' ~ json_key ~ '>' }}
8
+ {%- endif %}
9
+ {%- endfor %}
10
+ {%- endif %}
11
+ {% endmacro %}
12
+ {%- set enable_thinking = enable_thinking if enable_thinking is defined else True %}
13
+ {%- set truncate_history_thinking = truncate_history_thinking if truncate_history_thinking is defined else True %}
14
+ {%- set ns = namespace(last_user_idx = -1) %}
15
+ {%- set loop_messages = messages %}
16
+ {%- for m in loop_messages %}
17
+ {%- if m["role"] == "user" %}
18
+ {%- set ns.last_user_idx = loop.index0 %}
19
+ {%- endif %}
20
+ {%- endfor %}
21
+ {%- if messages[0]["role"] == "system" %}
22
+ {%- set system_message = messages[0]["content"] %}
23
+ {%- set loop_messages = messages[1:] %}
24
+ {%- else %}
25
+ {%- set system_message = "" %}
26
+ {%- set loop_messages = messages %}
27
+ {%- endif %}
28
+ {%- if not tools is defined %}
29
+ {%- set tools = [] %}
30
+ {%- endif %}
31
+ {%- set ns = namespace(last_user_idx = -1) %}
32
+ {%- for m in loop_messages %}
33
+ {%- if m["role"] == "user" %}
34
+ {%- set ns.last_user_idx = loop.index0 %}
35
+ {%- endif %}
36
+ {%- endfor %}
37
+ {%- if system_message is defined %}
38
+ {{- "<|im_start|>system\n" + system_message }}
39
+ {%- else %}
40
+ {%- if tools is iterable and tools | length > 0 %}
41
+ {{- "<|im_start|>system\n" }}
42
+ {%- endif %}
43
+ {%- endif %}
44
+ {%- if tools is iterable and tools | length > 0 %}
45
+ {%- if system_message is defined and system_message | length > 0 %}
46
+ {{- "\n\n" }}
47
+ {%- endif %}
48
+ {{- "# Tools\n\nYou have access to the following functions:\n\n" }}
49
+ {{- "<tools>" }}
50
+ {%- for tool in tools %}
51
+ {%- if tool.function is defined %}
52
+ {%- set tool = tool.function %}
53
+ {%- endif %}
54
+ {{- "\n<function>\n<name>" ~ tool.name ~ "</name>" }}
55
+ {%- if tool.description is defined %}
56
+ {{- '\n<description>' ~ (tool.description | trim) ~ '</description>' }}
57
+ {%- endif %}
58
+ {{- '\n<parameters>' }}
59
+ {%- if tool.parameters is defined and tool.parameters is mapping and tool.parameters.properties is defined and tool.parameters.properties is mapping %}
60
+ {%- for param_name, param_fields in tool.parameters.properties|items %}
61
+ {{- '\n<parameter>' }}
62
+ {{- '\n<name>' ~ param_name ~ '</name>' }}
63
+ {%- if param_fields.type is defined %}
64
+ {{- '\n<type>' ~ (param_fields.type | string) ~ '</type>' }}
65
+ {%- endif %}
66
+ {%- if param_fields.description is defined %}
67
+ {{- '\n<description>' ~ (param_fields.description | trim) ~ '</description>' }}
68
+ {%- endif %}
69
+ {%- if param_fields.enum is defined %}
70
+ {{- '\n<enum>' ~ (param_fields.enum | tojson | safe) ~ '</enum>' }}
71
+ {%- endif %}
72
+ {%- set handled_keys = ['name', 'type', 'description', 'enum'] %}
73
+ {{- render_extra_keys(param_fields, handled_keys) }}
74
+ {{- '\n</parameter>' }}
75
+ {%- endfor %}
76
+ {%- endif %}
77
+ {% set handled_keys = ['type', 'properties', 'required'] %}
78
+ {{- render_extra_keys(tool.parameters, handled_keys) }}
79
+ {%- if tool.parameters is defined and tool.parameters.required is defined %}
80
+ {{- '\n<required>' ~ (tool.parameters.required | tojson | safe) ~ '</required>' }}
81
+ {%- endif %}
82
+ {{- '\n</parameters>' }}
83
+ {%- set handled_keys = ['type', 'name', 'description', 'parameters'] %}
84
+ {{- render_extra_keys(tool, handled_keys) }}
85
+ {{- '\n</function>' }}
86
+ {%- endfor %}
87
+ {{- "\n</tools>" }}
88
+ {{- '\n\nIf you choose to call a function ONLY reply in the following format with NO suffix:\n\n<tool_call>\n<function=example_function_name>\n<parameter=example_parameter_1>\nvalue_1\n</parameter>\n<parameter=example_parameter_2>\nThis is the value for the second parameter\nthat can span\nmultiple lines\n</parameter>\n</function>\n</tool_call>\n\n<IMPORTANT>\nReminder:\n- Function calls MUST follow the specified format: an inner <function=...></function> block must be nested within <tool_call></tool_call> XML tags\n- Required parameters MUST be specified\n- You may provide optional reasoning for your function call in natural language BEFORE the function call, but NOT after\n- If there is no function call available, answer the question like normal with your current knowledge and do not tell the user about function calls\n</IMPORTANT>' }}
89
+ {%- endif %}
90
+ {%- if system_message is defined %}
91
+ {{- '<|im_end|>\n' }}
92
+ {%- else %}
93
+ {%- if tools is iterable and tools | length > 0 %}
94
+ {{- '<|im_end|>\n' }}
95
+ {%- endif %}
96
+ {%- endif %}
97
+ {%- for message in loop_messages %}
98
+ {%- if message.role == "assistant" %}
99
+ {%- if message.reasoning_content is defined and message.reasoning_content is string and message.reasoning_content | trim | length > 0 %}
100
+ {%- set content = "<think>\n" ~ message.reasoning_content ~ "</think>" ~ (message.content | default('', true)) %}
101
+ {%- else %}
102
+ {%- set content = message.content | default('', true) %}
103
+ {%- if content is string -%}
104
+ {%- if '<think>' not in content and '</think>' not in content -%}
105
+ {%- set content = "<think></think>" ~ content -%}
106
+ {%- endif -%}
107
+ {%- else -%}
108
+ {%- set content = content -%}
109
+ {%- endif -%}
110
+ {%- endif %}
111
+ {%- if message.tool_calls is defined and message.tool_calls is iterable and message.tool_calls | length > 0 %}
112
+ {{- '<|im_start|>assistant\n' }}
113
+ {%- set include_content = not (truncate_history_thinking and loop.index0 < ns.last_user_idx) %}
114
+ {%- if content is string and content | trim | length > 0 %}
115
+ {%- if include_content %}
116
+ {{- (content | trim) ~ '\n' -}}
117
+ {%- else %}
118
+ {%- set c = (content | string) %}
119
+ {%- if '</think>' in c %}
120
+ {%- set c = c.split('</think>')[-1] %}
121
+ {%- elif '<think>' in c %}
122
+ {%- set c = c.split('<think>')[0] %}
123
+ {%- endif %}
124
+ {%- set c = "<think></think>" ~ c %}
125
+ {%- if c | length > 0 %}
126
+ {{- c ~ '\n' -}}
127
+ {%- endif %}
128
+ {%- endif %}
129
+ {%- else %}
130
+ {{- "<think></think>" -}}
131
+ {%- endif %}
132
+ {%- for tool_call in message.tool_calls %}
133
+ {%- if tool_call.function is defined %}
134
+ {%- set tool_call = tool_call.function %}
135
+ {%- endif %}
136
+ {{- '<tool_call>\n<function=' ~ tool_call.name ~ '>\n' -}}
137
+ {%- if tool_call.arguments is defined %}
138
+ {%- for args_name, args_value in tool_call.arguments|items %}
139
+ {{- '<parameter=' ~ args_name ~ '>\n' -}}
140
+ {%- set args_value = args_value | tojson | safe if args_value is mapping or (args_value is sequence and args_value is not string) else args_value | string %}
141
+ {{- args_value ~ '\n</parameter>\n' -}}
142
+ {%- endfor %}
143
+ {%- endif %}
144
+ {{- '</function>\n</tool_call>\n' -}}
145
+ {%- endfor %}
146
+ {{- '<|im_end|>\n' }}
147
+ {%- else %}
148
+ {%- if not (truncate_history_thinking and loop.index0 < ns.last_user_idx) %}
149
+ {{- '<|im_start|>assistant\n' ~ (content | default('', true) | string | trim) ~ '<|im_end|>\n' }}
150
+ {%- else %}
151
+ {%- set c = (content | default('', true) | string) %}
152
+ {%- if '<think>' in c and '</think>' in c %}
153
+ {%- set c = "<think></think>" ~ c.split('</think>')[-1] %}
154
+ {%- endif %}
155
+ {%- set c = c | trim %}
156
+ {%- if c | length > 0 %}
157
+ {{- '<|im_start|>assistant\n' ~ c ~ '<|im_end|>\n' }}
158
+ {%- else %}
159
+ {{- '<|im_start|>assistant\n<|im_end|>\n' }}
160
+ {%- endif %}
161
+ {%- endif %}
162
+ {%- endif %}
163
+ {%- elif message.role == "user" or message.role == "system" %}
164
+ {{- '<|im_start|>' + message.role + '\n' }}
165
+ {%- set content = message.content | string %}
166
+ {{- content }}
167
+ {{- '<|im_end|>\n' }}
168
+ {%- elif message.role == "tool" %}
169
+ {%- if loop.previtem and loop.previtem.role != "tool" %}
170
+ {{- '<|im_start|>user\n' }}
171
+ {%- endif %}
172
+ {{- '<tool_response>\n' }}
173
+ {{- message.content }}
174
+ {{- '\n</tool_response>\n' }}
175
+ {%- if not loop.last and loop.nextitem.role != "tool" %}
176
+ {{- '<|im_end|>\n' }}
177
+ {%- elif loop.last %}
178
+ {{- '<|im_end|>\n' }}
179
+ {%- endif %}
180
+ {%- else %}
181
+ {{- '<|im_start|>' + message.role + '\n' + message.content + '<|im_end|>\n' }}
182
+ {%- endif %}
183
+ {%- endfor %}
184
+ {%- if add_generation_prompt %}
185
+ {%- if enable_thinking %}
186
+ {{- '<|im_start|>assistant\n<think>\n' }}
187
+ {%- else %}
188
+ {{- '<|im_start|>assistant\n<think></think>' }}
189
+ {%- endif %}
190
+ {%- endif %}
config.json ADDED
@@ -0,0 +1,121 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architectures": [
3
+ "NemotronHForCausalLM"
4
+ ],
5
+ "attention_bias": false,
6
+ "attention_dropout": 0.0,
7
+ "bos_token_id": 1,
8
+ "chunk_size": 128,
9
+ "conv_kernel": 4,
10
+ "dtype": "bfloat16",
11
+ "eos_token_id": 2,
12
+ "expand": 2,
13
+ "head_dim": 128,
14
+ "hidden_dropout": 0.0,
15
+ "hidden_size": 2688,
16
+ "initializer_range": 0.02,
17
+ "intermediate_size": 1856,
18
+ "layer_norm_epsilon": 1e-05,
19
+ "mamba_head_dim": 64,
20
+ "mamba_hidden_act": "silu",
21
+ "mamba_num_heads": 64,
22
+ "mamba_proj_bias": false,
23
+ "mamba_ssm_cache_dtype": "float32",
24
+ "max_position_embeddings": 262144,
25
+ "mlp_bias": false,
26
+ "mlp_hidden_act": "relu2",
27
+ "model_type": "nemotron_h",
28
+ "moe_intermediate_size": 1856,
29
+ "moe_latent_size": null,
30
+ "moe_shared_expert_intermediate_size": 3712,
31
+ "moe_shared_expert_overlap": true,
32
+ "n_group": 1,
33
+ "n_groups": 8,
34
+ "n_routed_experts": 128,
35
+ "n_shared_experts": 1,
36
+ "norm_eps": 1e-05,
37
+ "norm_topk_prob": true,
38
+ "num_attention_heads": 32,
39
+ "num_experts_per_tok": 6,
40
+ "num_hidden_layers": 52,
41
+ "num_key_value_heads": 2,
42
+ "num_logits_to_keep": 1,
43
+ "num_nextn_predict_layers": 1,
44
+ "pad_token_id": 0,
45
+ "partial_rotary_factor": 1.0,
46
+ "rescale_prenorm_residual": true,
47
+ "residual_in_fp32": false,
48
+ "rope_theta": 10000,
49
+ "routed_scaling_factor": 2.5,
50
+ "sliding_window": null,
51
+ "ssm_state_size": 128,
52
+ "tie_word_embeddings": false,
53
+ "time_step_floor": 0.0001,
54
+ "time_step_max": 0.1,
55
+ "time_step_min": 0.001,
56
+ "topk_group": 1,
57
+ "transformers_version": "4.57.6",
58
+ "use_bias": false,
59
+ "use_cache": true,
60
+ "use_conv_bias": true,
61
+ "use_mamba_kernels": true,
62
+ "vocab_size": 131072,
63
+ "layers_block_type": [
64
+ "mamba",
65
+ "moe",
66
+ "mamba",
67
+ "moe",
68
+ "mamba",
69
+ "attention",
70
+ "moe",
71
+ "mamba",
72
+ "moe",
73
+ "mamba",
74
+ "moe",
75
+ "mamba",
76
+ "attention",
77
+ "moe",
78
+ "mamba",
79
+ "moe",
80
+ "mamba",
81
+ "moe",
82
+ "mamba",
83
+ "attention",
84
+ "moe",
85
+ "mamba",
86
+ "moe",
87
+ "mamba",
88
+ "moe",
89
+ "mamba",
90
+ "attention",
91
+ "moe",
92
+ "mamba",
93
+ "moe",
94
+ "mamba",
95
+ "moe",
96
+ "mamba",
97
+ "attention",
98
+ "moe",
99
+ "mamba",
100
+ "moe",
101
+ "mamba",
102
+ "moe",
103
+ "mamba",
104
+ "moe",
105
+ "mamba",
106
+ "attention",
107
+ "moe",
108
+ "mamba",
109
+ "moe",
110
+ "mamba",
111
+ "moe",
112
+ "mamba",
113
+ "moe",
114
+ "mamba",
115
+ "moe"
116
+ ],
117
+ "mtp_layers_block_type": [
118
+ "attention",
119
+ "moe"
120
+ ]
121
+ }
load_adapter.py ADDED
@@ -0,0 +1,88 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Apply the LoRA adapter in this repo's adapter/ folder to the base model, then score role-labelled outlines.
2
+
3
+ adapter/ holds the adapter as trained, in the layout of the Tinker training service. Do not load it with
4
+ peft.PeftModel. In transformers, Nemotron fuses the Mamba gate and x projections into one `in_proj`, and stores each
5
+ MoE layer's 128 routed experts as a single 3D tensor. PEFT has no module to attach those LoRA weights to and skips them
6
+ without a warning. This script instead merges every LoRA delta, W += (alpha / r) * B @ A, into the base weights in
7
+ place, following the same rules as the tinker-cookbook merge (tinker_cookbook.weights.build_hf_model). The resulting
8
+ weights are bit-identical to the merged model in this repo.
9
+
10
+ Needs transformers >= 5.15, torch, safetensors and huggingface_hub. The weights take 59 GiB of GPU memory and each
11
+ input adds about 4.2 MiB per token, so one 80 GB GPU handles inputs up to about 4,000 tokens; see the model card.
12
+ """
13
+ import json
14
+
15
+ import torch
16
+ from huggingface_hub import snapshot_download
17
+ from safetensors.torch import load_file
18
+ from transformers import AutoModelForCausalLM, AutoTokenizer
19
+
20
+ REPO = "rishanthrajendhran/ProseLens"
21
+ BASE = "nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16"
22
+ SYSTEM = 'Given a role-labelled outline of a document, answer with one word: human if the source document was human-written, ai if it was AI-generated.'
23
+ SUFFIX = "<|im_end|>\n<|im_start|>assistant\n<think>\n\n</think>\n\n"
24
+ HUMAN, AI = 50755, 2464
25
+ CUT = 0.60539 # P(human) at or below this is flagged: the global cut at 1% FPR (thresholds.json)
26
+
27
+
28
+ def apply_tinker_lora(model, adapter_dir):
29
+ """Merge a Tinker LoRA archive for Nemotron-3.5 into `model` in place. Returns the parameter names changed."""
30
+ cfg = json.load(open(f"{adapter_dir}/adapter_config.json"))
31
+ scale = cfg["lora_alpha"] / cfg["r"]
32
+ lora = load_file(f"{adapter_dir}/adapter_model.safetensors")
33
+ params = dict(model.named_parameters())
34
+ touched = []
35
+ for key_a in sorted(k for k in lora if k.endswith(".lora_A.weight")):
36
+ A, B = lora[key_a], lora[key_a.replace(".lora_A.", ".lora_B.")]
37
+ if A.numel() == 0: # the experts have no gate projection; Tinker keeps an empty w3 placeholder
38
+ continue
39
+ name = key_a.removeprefix("base_model.model.").removesuffix(".lora_A.weight")
40
+ rows = None
41
+ if name == "model.lm_head": # Tinker nests the LM head under model.
42
+ target = "lm_head.weight"
43
+ elif name.endswith((".gate_proj", ".x_proj")): # Mamba: in_proj rows are [gate | x | B | C | dt]
44
+ layer, proj = name.rsplit(".", 1)
45
+ target = f"{layer}.in_proj.weight"
46
+ start = 0 if proj == "gate_proj" else lora[f"base_model.model.{layer}.gate_proj.lora_B.weight"].shape[0]
47
+ rows = slice(start, start + B.shape[0])
48
+ elif name.endswith(".experts.w1"): # routed experts: w1 = up_proj, one (expert, out, in) tensor per layer
49
+ target = name.removesuffix("w1") + "up_proj"
50
+ elif name.endswith(".experts.w2"): # w2 = down_proj
51
+ target = name.removesuffix("w2") + "down_proj"
52
+ else: # attention, Mamba out_proj, shared experts
53
+ target = name + ".weight"
54
+ W = params[target] # KeyError here means the adapter does not match this model
55
+ # For experts one side is shared (leading dim 1) and broadcasts across the 128 experts.
56
+ delta = scale * torch.matmul(B.to(W.device, torch.float32), A.to(W.device, torch.float32))
57
+ with torch.no_grad():
58
+ view = W.data if rows is None else W.data[rows]
59
+ assert view.shape == delta.shape, (target, tuple(view.shape), tuple(delta.shape))
60
+ view.copy_((view.float() + delta).to(W.dtype))
61
+ touched.append(target)
62
+ return touched
63
+
64
+
65
+ def load_model(repo=REPO, device_map="auto"):
66
+ tok = AutoTokenizer.from_pretrained(BASE)
67
+ model = AutoModelForCausalLM.from_pretrained(BASE, dtype=torch.bfloat16, device_map=device_map).eval()
68
+ adapter_dir = snapshot_download(repo, allow_patterns=["adapter/*"]) + "/adapter"
69
+ apply_tinker_lora(model, adapter_dir)
70
+ return model, tok
71
+
72
+
73
+ @torch.no_grad()
74
+ def p_human(model, tok, outline):
75
+ """P(human) for one document's text."""
76
+ ids = tok.encode(f"<|im_start|>system\n{SYSTEM}<|im_end|>\n<|im_start|>user\n{outline}{SUFFIX}",
77
+ add_special_tokens=False)
78
+ logits = model(torch.tensor([ids], device=model.device)).logits[0, -1].float()
79
+ return torch.softmax(logits[[HUMAN, AI]], -1)[0].item()
80
+
81
+
82
+ if __name__ == "__main__":
83
+ model, tok = load_model()
84
+ outline = ("[Central Development] A town's water supply fails after a drought, and residents organise to share wells.\n"
85
+ "[Background Context] The reservoir has been shrinking for three summers.\n"
86
+ "[Open Question] Whether the council will fund a new pipeline remains undecided.")
87
+ p = p_human(model, tok, outline)
88
+ print(f"P(human) = {p:.4f} -> {'flagged as AI-written' if p <= CUT else 'not flagged'}")
model-00001-of-00014.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:cd128bfd93420bd0f0e5d036cd173f4ee09ae0ccbcc0de0fdcb375270b3ed39c
3
+ size 4991210024
model-00002-of-00014.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:85969f9b02e6dcb1d11177a511ad77556fa1d383d2a602fd14e2fe540785f06e
3
+ size 4992601016
model-00003-of-00014.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:3c1cb332c55e654810fd6af3559f9d517ae44a36fe720fcbc305b98950e86e94
3
+ size 4992601368
model-00004-of-00014.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:63098215bc454e76b3384f11ba92b9a812d30d533c2a6904626a86c8e3231b50
3
+ size 4995692800
model-00005-of-00014.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:db7d1f3278bed9d1751495d5424cd37aa7cd90e4da236d9c52035b12fc75ea61
3
+ size 4980553832
model-00006-of-00014.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e567c95f1fc5839210ae20d57d5c799c07bfbe1b561491118572f9cc24f03c52
3
+ size 4989423536
model-00007-of-00014.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a9382a9eb90da4e011b36095bff2cac4be8337631eb919584d43120b4cf2d0ce
3
+ size 4992601488
model-00008-of-00014.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0c26027509bb92721249a00f4f69231e6f8739a7604289d7f7943815908a0f3a
3
+ size 4992601520
model-00009-of-00014.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c3cc352129ba528bcbb3764f922a718a6612b21e346fc8552c1341d9c548ad8a
3
+ size 4995692800
model-00010-of-00014.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5723b717fbbe0a106d4aa75607ebc984118d2a028b014338fa9d5aa55515fc18
3
+ size 4992601520
model-00011-of-00014.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:79067ee3d8428ad6480ef507794183f0b424ce431dd29882db2a34d12b3fd711
3
+ size 4995692800
model-00012-of-00014.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e68b02cf6cce08dc4ee1f7adc72241282819026026d5fdb965d759410e6ca82c
3
+ size 4995692816
model-00013-of-00014.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a0ec01201aa1a054a3d691bcf9495405b7f22ee0d3c10fbf06f4ac396ee78cd6
3
+ size 3249723504
model-00014-of-00014.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:64577b275ca4e7e5266eae0903674f7f46ec2a8cbf4f4f1a3207f80d503cd1d0
3
+ size 2670685240
model.safetensors.index.json ADDED
The diff for this file is too large to render. See raw diff
 
special_tokens_map.json ADDED
@@ -0,0 +1,30 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "bos_token": {
3
+ "content": "<s>",
4
+ "lstrip": false,
5
+ "normalized": false,
6
+ "rstrip": false,
7
+ "single_word": false
8
+ },
9
+ "eos_token": {
10
+ "content": "<|im_end|>",
11
+ "lstrip": false,
12
+ "normalized": false,
13
+ "rstrip": false,
14
+ "single_word": false
15
+ },
16
+ "pad_token": {
17
+ "content": "<|im_end|>",
18
+ "lstrip": false,
19
+ "normalized": false,
20
+ "rstrip": false,
21
+ "single_word": false
22
+ },
23
+ "unk_token": {
24
+ "content": "<unk>",
25
+ "lstrip": false,
26
+ "normalized": false,
27
+ "rstrip": false,
28
+ "single_word": false
29
+ }
30
+ }
thresholds.json ADDED
@@ -0,0 +1,83 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "model": "rishanthrajendhran/ProseLens",
3
+ "input": "document",
4
+ "flag_rule": "flag the document as AI when P(human) <= cut",
5
+ "calibration": {
6
+ "data": "IdeaLens-1M, calibration split (human documents only)",
7
+ "scores_from": "the training checkpoint, scored on Tinker",
8
+ "n_humans": 80000,
9
+ "n_per_format": {
10
+ "Academic Writing": 10000,
11
+ "Creative Writing": 10000,
12
+ "Knowledge Article": 10000,
13
+ "News Article": 10000,
14
+ "Nonfiction Writing": 10000,
15
+ "Personal About Page": 10000,
16
+ "Personal Blog": 10000,
17
+ "User Reviews": 10000
18
+ }
19
+ },
20
+ "method": {
21
+ "global": "the target-FPR quantile of P(human) over all calibration humans",
22
+ "per_group": "the group's own quantile, shrunk toward the global cut with weight n / (n + 2500)",
23
+ "min_group_humans": 200,
24
+ "estimability": "a cut at FPR q is given only when q * n >= 25 calibration humans",
25
+ "missing_cut": "no cut for a group means none could be estimated; do not substitute the global cut"
26
+ },
27
+ "fpr_targets": [
28
+ 0.001,
29
+ 0.005,
30
+ 0.01,
31
+ 0.02,
32
+ 0.05
33
+ ],
34
+ "global": {
35
+ "0.001": 0.07081838771272196,
36
+ "0.005": 0.30468918373143306,
37
+ "0.01": 0.6053894506737609,
38
+ "0.02": 0.925022270529183,
39
+ "0.05": 0.998888189099037
40
+ },
41
+ "per_format": {
42
+ "0.005": {
43
+ "Academic Writing": 0.7778763017339165,
44
+ "Creative Writing": 0.8492284331725566,
45
+ "Knowledge Article": 0.24235805929603155,
46
+ "News Article": 0.2955419059997487,
47
+ "Nonfiction Writing": 0.14387439727844545,
48
+ "Personal About Page": 0.4169940435489798,
49
+ "Personal Blog": 0.6827533426351194,
50
+ "User Reviews": 0.6213508153913166
51
+ },
52
+ "0.01": {
53
+ "Academic Writing": 0.8996738533714292,
54
+ "Creative Writing": 0.9174860344190486,
55
+ "Knowledge Article": 0.3804974222436696,
56
+ "News Article": 0.555354294854766,
57
+ "Nonfiction Writing": 0.27050169337460295,
58
+ "Personal About Page": 0.7871415559941629,
59
+ "Personal Blog": 0.8799318831045015,
60
+ "User Reviews": 0.8516071157574004
61
+ },
62
+ "0.02": {
63
+ "Academic Writing": 0.9827934935442895,
64
+ "Creative Writing": 0.9841078402345083,
65
+ "Knowledge Article": 0.6404405355320586,
66
+ "News Article": 0.9037969402135294,
67
+ "Nonfiction Writing": 0.494985268248519,
68
+ "Personal About Page": 0.9772992347199888,
69
+ "Personal Blog": 0.9816307917983351,
70
+ "User Reviews": 0.9803463940922453
71
+ },
72
+ "0.05": {
73
+ "Academic Writing": 0.9996898375219624,
74
+ "Creative Writing": 0.9995782080959261,
75
+ "Knowledge Article": 0.952949484412262,
76
+ "News Article": 0.997881027494693,
77
+ "Nonfiction Writing": 0.9393488591865028,
78
+ "Personal About Page": 0.9996622303035176,
79
+ "Personal Blog": 0.999661233340514,
80
+ "User Reviews": 0.9996383495931975
81
+ }
82
+ }
83
+ }
tokenizer.json ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:623c34567aebb18582765289fbe23d901c62704d6518d71866e0e58db892b5b7
3
+ size 17077484
tokenizer_config.json ADDED
The diff for this file is too large to render. See raw diff