ehsantorabi commited on
Commit
d9649b7
·
verified ·
1 Parent(s): da6cd20

Upload README.md with huggingface_hub

Browse files
Files changed (1) hide show
  1. README.md +97 -0
README.md ADDED
@@ -0,0 +1,97 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: mit
3
+ language:
4
+ - fa
5
+ library_name: pytorch
6
+ pipeline_tag: text-classification
7
+ tags:
8
+ - ai-generated-text-detection
9
+ - persian
10
+ - farsi
11
+ - text-classification
12
+ - stylometry
13
+ - ensemble
14
+ - bilstm
15
+ - textcnn
16
+ - transformer
17
+ - parsbert
18
+ - gpt2
19
+ - laya
20
+ ---
21
+
22
+ # AITextD — Persian AI-Generated Text Detection Models
23
+
24
+ Trained checkpoints for the [AITextDetection](https://github.com/ehsan-torabi/AITextDetection)
25
+ project: detecting whether **Persian** text was written by a human or generated by AI.
26
+
27
+ This repository mirrors the `models/` directory of the source project one-to-one, so a
28
+ single download drops every serving artifact into place:
29
+
30
+ ```bash
31
+ # from the project root, after cloning the GitHub repo
32
+ hf download ehsantorabi/aitextd-models --local-dir models/
33
+ ```
34
+
35
+ ## What is inside
36
+
37
+ Sixteen registered models cover four families — classical ML on TF-IDF (+ stylometric
38
+ features), deep networks trained from scratch, fine-tuned LLMs, and a System One
39
+ decision model — plus the shared serving artifacts (`tfidf_vectorizer.joblib`, the
40
+ GPT-2 and ParsBERT tokenizers, the raw Persian GPT-2 LM used for the training-free
41
+ perplexity signal, and the fine-tuned Laya decision model). The weighted **ensemble**
42
+ is computed live by the serving stack from these checkpoints; it is not a file.
43
+
44
+ | Model | id | Artifact | Architecture | Family | Ensemble weight | Test accuracy | Test F1 |
45
+ |---|---|---|---|---|---|---|---|
46
+ | BiLSTM | `bilstm` | `lstmModel.pt` | BiLSTM | Deep learning | 0.7 | 0.9832 | 0.9831 |
47
+ | TextCNN | `textcnn` | `textcnnModel.pt` | TextCNN | Deep learning | 0.5 | 0.9905 | 0.9905 |
48
+ | TextCNN (RLCD fine-tune) | `textcnn_rlcd` | `textcnnRlcdModel.pt` | TextCNN + RLCD fine-tune | Deep learning | 0.0 | 0.9900 | 0.9900 |
49
+ | Attention BiLSTM | `attention_bilstm` | `attnBilstmModel.pt` | BiLSTM + additive attention pooling | Deep learning | 0.5 | 0.9881 | 0.9881 |
50
+ | GPT-2 Transformer (fine-tuned) | `gpt2_transformer` | `gpt2ClassifierModel.pt` | GPT2ForSequenceClassification (full fine-tune) | Deep learning | 3.0 | 0.9969 | 0.9969 |
51
+ | ParsBERT (Bidirectional Encoder) | `parsbert` | `parsbertClassifierModel.pt` | BertModel + linear head on [CLS] | Deep learning | 3.0 | 0.9983 | 0.9983 |
52
+ | System One (Decision Head) | `system_one` | `systemOneModel.pt` | ParsBERT (fine-tuned encoder) + typed decision heads (choice/score/noul) | Deep learning | 0.6 | 0.9893 | 0.9893 |
53
+ | SVM | `svm` | `svModel.joblib` | LinearSVC | Classical ML | 1.0 | 0.9967 | 0.9967 |
54
+ | Logistic Regression | `logreg` | `logregModel.joblib` | LogisticRegression | Classical ML | 1.0 | 0.9964 | 0.9964 |
55
+ | Naive Bayes | `naive_bayes` | `nbModel.joblib` | MultinomialNB | Classical ML | 1.0 | 0.9810 | 0.9809 |
56
+ | Random Forest | `random_forest` | `rfModel.joblib` | RandomForestClassifier | Classical ML | 1.0 | 0.9886 | 0.9886 |
57
+ | Gradient Boosting | `gradient_boosting` | `gbModel.joblib` | GradientBoostingClassifier | Classical ML | 1.0 | 0.9912 | 0.9912 |
58
+ | Laya (System One, fine-tuned) | `laya` | `laya_persian` (directory) | Laya multilingual (mmBERT-base) + typed decision heads (RLCD fine-tune) | Deep learning | 2.0 | 0.9993 | 0.9993 |
59
+ | MLP | `mlp` | `mlpModel.pt` | MLP (2 hidden layers) | Deep learning | 0.5 | 0.9964 | 0.9965 |
60
+ | CharCNN | `charcnn` | `charcnnModel.pt` | CharCNN (character-level TextCNN) | Deep learning | 0.5 | 0.9983 | 0.9983 |
61
+ | Tiny Transformer (from scratch) | `tinytransformer` | `tinyTransformerModel.pt` | Transformer encoder (2 layers, from scratch) | Deep learning | 0.5 | 0.9891 | 0.9891 |
62
+
63
+ ## Artifacts not included
64
+
65
+ - `models/laya_multilingual/` — the untouched
66
+ [convaiinnovations/laya](https://huggingface.co/convaiinnovations/laya) *multilingual*
67
+ (mmBERT-base) base checkpoint; only needed to re-run the fine-tuning notebook, so it is
68
+ not duplicated here. Serving uses the fine-tuned `laya_persian/` above.
69
+ - `models/laya_persian/checkpoint_latest/` — a redundant mid-training checkpoint of
70
+ `laya_persian` (identical final weights); skipped.
71
+
72
+ ## Evaluation protocol
73
+
74
+ Metrics above are test-set results from the project's leak-free v2 protocol
75
+ (group-aware splits, paraphrase-attack and domain-shift conditions). Frozen evidence,
76
+ per-model confusion matrices and full tables live in the GitHub repo under `reports/`
77
+ and `models.json`.
78
+
79
+ ## Base-model attribution
80
+
81
+ - ParsBERT classifier → [HooshvareLab/bert-fa-base-uncased](https://huggingface.co/HooshvareLab/bert-fa-base-uncased)
82
+ - GPT-2 classifier + perplexity LM → [HooshvareLab/gpt2-fa](https://huggingface.co/HooshvareLab/gpt2-fa)
83
+ - Laya System One decision model → [convaiinnovations/laya](https://huggingface.co/convaiinnovations/laya) (multilingual, Apache-2.0)
84
+
85
+ ## Usage
86
+
87
+ The checkpoints are consumed by the project's FastAPI serving stack (`aitextd/`,
88
+ `models.json`), not by a single `pipeline(...)`. See the
89
+ [GitHub repo](https://github.com/ehsan-torabi/AITextDetection) for the serving API,
90
+ the six training notebooks, the web front end and tests:
91
+
92
+ ```bash
93
+ uv sync
94
+ hf download ehsantorabi/aitextd-models --local-dir models/
95
+ uv run python server.py # API on :8000
96
+ ```
97
+