Instructions to use deepquillapp/trait-verifier with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Laya
How to use deepquillapp/trait-verifier with Laya:
# No code snippets available yet for this library. # To use this model, check the repository files and the library's documentation. # Want to help? PRs adding snippets are welcome at: # https://github.com/huggingface/huggingface.js
- Notebooks
- Google Colab
- Kaggle
Upload folder using huggingface_hub
Browse files- README.md +44 -0
- model.w8.onnx +3 -0
- tokenizer.json +0 -0
- 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 |
+
}
|