partnerinflight commited on
Commit
ec38dbd
·
verified ·
1 Parent(s): 43802f6

Upload folder using huggingface_hub

Browse files
Files changed (4) hide show
  1. README.md +44 -0
  2. model.w8.onnx +3 -0
  3. tokenizer.json +0 -0
  4. verifier_config.json +88 -0
README.md ADDED
@@ -0,0 +1,44 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: apache-2.0
3
+ base_model: convaiinnovations/laya
4
+ library_name: onnxruntime
5
+ pipeline_tag: text-classification
6
+ tags: [typed-decisions, laya, modernbert, fiction, character-traits]
7
+ ---
8
+
9
+ # DeepQuill trait verifier
10
+
11
+ A fine-tune of [Laya](https://huggingface.co/convaiinnovations/laya) (ModernBERT-large encoder plus a typed-decision head). [DeepQuill](https://deepquill.app) uses it to check character-trait claims that its local extractor pulls from fiction.
12
+
13
+ For each claim (a target character, `field = value`, the quote, and the passage), the model answers four yes/no questions:
14
+
15
+ - **aboutTarget**: is it about that character?
16
+ - **lasting**: a lasting attribute, not a momentary state?
17
+ - **literal**: literal, not figurative?
18
+ - **asserted**: asserted by the narration?
19
+
20
+ The claim is kept when all four pass.
21
+
22
+ ## Files
23
+
24
+ - `model.w8.onnx`: weight-only 8-bit (MatMulNBits). Activations run in fp32, so a claim's result doesn't depend on what it's batched with.
25
+ - `tokenizer.json`
26
+ - `verifier_config.json`: the exact question texts, sequence limits and calibration temperature.
27
+
28
+ ## Training data
29
+
30
+ Everything the model was trained on is public:
31
+
32
+ - 3,760 labelled claims from real extractor output over US-public-domain novels from 1894–1930 (Project Gutenberg).
33
+ - Synthetic contemporary passages written for this purpose.
34
+
35
+ The model was never trained on user manuscripts.
36
+
37
+ ## Evaluation
38
+
39
+ The model was scored on held-out sets that share no books or characters with the training data. Metric: keep-F1.
40
+
41
+ | Test set | This model | Qwen3-4B LLM verifier | No verifier |
42
+ |---|---|---|---|
43
+ | 1920s public-domain claims (524) | 0.75 | 0.32 | 0.45 |
44
+ | Contemporary manuscript claims (157) | 0.88 | 0.24 | 0.45 |
model.w8.onnx ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b805413395dbfa1c204eb5b542a5c658c5e249ac4ca70b63edc2714d7c4aed32
3
+ size 599059203
tokenizer.json ADDED
The diff for this file is too large to render. See raw diff
 
verifier_config.json ADDED
@@ -0,0 +1,88 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "schema": "deepquill-laya-verifier-v1",
3
+ "checks": [
4
+ "aboutTarget",
5
+ "lasting",
6
+ "literal",
7
+ "asserted"
8
+ ],
9
+ "physical_fields": [
10
+ "age",
11
+ "build",
12
+ "distinguishingFeatures",
13
+ "eyeColor",
14
+ "hairColor",
15
+ "hairStyle",
16
+ "height",
17
+ "skinTone"
18
+ ],
19
+ "questions": {
20
+ "physical": {
21
+ "aboutTarget": {
22
+ "ins": "The quote describes the TARGET character themselves.",
23
+ "true": "it is about the TARGET (pronouns resolved by the passage)",
24
+ "false": "it is about someone else, a portrait, or a nearby object"
25
+ },
26
+ "lasting": {
27
+ "ins": "The claim is a lasting attribute of the TARGET's body.",
28
+ "true": "a lasting body attribute: build, height, age, eye or hair colour, a scar, a habitual complexion",
29
+ "false": "a momentary state (a flush, pallor from shock, an expression, hair wet or tumbled), clothing or props, or a skill, habit or disposition"
30
+ },
31
+ "literal": {
32
+ "ins": "The claim is a literal description.",
33
+ "true": "literal, including conventional descriptive idioms",
34
+ "false": "a metaphor, simile or figure of speech carries it"
35
+ },
36
+ "asserted": {
37
+ "ins": "The narration asserts the claim as fact.",
38
+ "true": "narrated as fact",
39
+ "false": "negated, hedged, hypothetical, imagined, hearsay, or only said in dialogue"
40
+ }
41
+ },
42
+ "disposition": {
43
+ "aboutTarget": {
44
+ "ins": "The quote describes the TARGET character themselves.",
45
+ "true": "it is about the TARGET (pronouns resolved by the passage)",
46
+ "false": "it is about someone else, a portrait, or a nearby object"
47
+ },
48
+ "lasting": {
49
+ "ins": "The claim is a disposition or want of the TARGET that persists beyond this scene.",
50
+ "true": "a lasting trait, goal or flaw",
51
+ "false": "a passing intention, feeling or single act"
52
+ },
53
+ "literal": {
54
+ "ins": "The claim is a literal description.",
55
+ "true": "literal, including conventional descriptive idioms",
56
+ "false": "a metaphor, simile or figure of speech carries it"
57
+ },
58
+ "asserted": {
59
+ "ins": "The narration asserts the claim as fact.",
60
+ "true": "narrated as fact",
61
+ "false": "negated, hedged, hypothetical, imagined, hearsay, or only said in dialogue"
62
+ }
63
+ }
64
+ },
65
+ "noul_labels": {
66
+ "false": "false",
67
+ "true": "true"
68
+ },
69
+ "state_keys": [
70
+ "target",
71
+ "claim",
72
+ "quote",
73
+ "passage"
74
+ ],
75
+ "max_len": 512,
76
+ "head_max_len": 192,
77
+ "option_max_tokens": 48,
78
+ "temperature": 1.2899174690246582,
79
+ "threshold": 0.5,
80
+ "inputs": [
81
+ "input_ids",
82
+ "attention_mask",
83
+ "marker_pos",
84
+ "marker_mask",
85
+ "qtype"
86
+ ],
87
+ "output": "logits"
88
+ }