crystas commited on
Commit
2a3dce9
·
verified ·
1 Parent(s): 7e7b191

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +139 -1
README.md CHANGED
@@ -16,4 +16,142 @@ widget:
16
  - text: search for previous sentence delete
17
  base_model:
18
  - distilbert/distilbert-base-uncased
19
- ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
16
  - text: search for previous sentence delete
17
  base_model:
18
  - distilbert/distilbert-base-uncased
19
+ ---
20
+
21
+
22
+ # distilbert-command-data-tagger
23
+
24
+ `distilbert-command-data-tagger` is a lightweight, optimized ONNX model trained for joint **Intent Classification** and **Named Entity Recognition (NER) / Data Tagging** from natural language user commands.
25
+
26
+ It is designed for lightweight deployment on low-resource runtimes and Hugging Face Inference Endpoints using ONNX Runtime.
27
+
28
+ ---
29
+
30
+ ## Model Architecture & Files
31
+
32
+ - **Base Model**: `distilbert-base-uncased`
33
+ - **Model Format**: ONNX (`joint_command_parser.onnx` + `joint_command_parser.onnx.data`)
34
+ - **Max Sequence Length**: 64 tokens
35
+ - **Output Heads**:
36
+ 1. **Intent Logits**: Decodes top-level command intent (with a strict `0.65` confidence threshold fallback to `unknown`).
37
+ 2. **NER / Data Tag Logits**: Tagging token sequences (e.g., target text to edit, recipient email address, or search queries).
38
+
39
+ ---
40
+
41
+ ## Supported Intents
42
+
43
+ The model categorizes inputs into one of the following structured categories:
44
+
45
+ - **Document Editing**: `add_text`, `search_delete`, `prev_sentence_delete`, `bold`, `italic`, `underline`
46
+ - **Navigation**: `navigate_email`, `navigate_dm`
47
+ - **Email Page Operations**: `email_compose`, `email_send`, `email_reply`, `email_forward`, `email_delete`, `email_search`
48
+ - **Direct Message (DM) Operations**: `dm_send`, `dm_reply`, `dm_delete`, `dm_search`
49
+ - **Fallback**: `unknown`
50
+
51
+ ---
52
+
53
+ ## Inference with Custom Endpoint Handler (`handler.py`)
54
+
55
+ Below is the standard Hugging Face `EndpointHandler` used to load and run inference on the ONNX graph:
56
+
57
+ ```python
58
+ import os
59
+ import numpy as np
60
+ import onnxruntime as ort
61
+ from transformers import AutoTokenizer
62
+
63
+ class EndpointHandler:
64
+ def __init__(self, path=""):
65
+ # 1. Resolve paths for BOTH structural graph and matrix weights
66
+ model_path = os.path.join(path, "joint_command_parser.onnx")
67
+
68
+ # Initialize the ONNX Runtime execution engine thread pool
69
+ self.session = ort.InferenceSession(model_path, providers=["CPUExecutionProvider"])
70
+
71
+ # 2. Match the exact tokenizer backbone used during training
72
+ self.tokenizer = AutoTokenizer.from_pretrained("distilbert-base-uncased")
73
+
74
+ # 3. Structural target layouts
75
+ self.intents = [
76
+ # document editing
77
+ "add_text",
78
+ "search_delete",
79
+ "prev_sentence_delete",
80
+ "bold",
81
+ "italic",
82
+ "underline",
83
+ # navigation
84
+ "navigate_email",
85
+ "navigate_dm",
86
+ # email page
87
+ "email_compose",
88
+ "email_send",
89
+ "email_reply",
90
+ "email_forward",
91
+ "email_delete",
92
+ "email_search",
93
+ # dm page
94
+ "dm_send",
95
+ "dm_reply",
96
+ "dm_delete",
97
+ "dm_search",
98
+ # fallback
99
+ "unknown",
100
+ ]
101
+ self.max_len = 64
102
+
103
+ def __call__(self, data):
104
+ # Extract input text payload sent via HTTP POST Request
105
+ inputs_payload = data.get("inputs", "")
106
+ if not inputs_payload:
107
+ return {"error": "Missing 'inputs' string parameter inside payload."}
108
+
109
+ # Tokenize incoming sequence matching matrix bounds
110
+ tokenized = self.tokenizer(
111
+ inputs_payload,
112
+ max_length=self.max_len,
113
+ padding="max_length",
114
+ truncation=True,
115
+ return_tensors="np"
116
+ )
117
+
118
+ onnx_feeds = {
119
+ "input_ids": tokenized["input_ids"].astype(np.int64),
120
+ "attention_mask": tokenized["attention_mask"].astype(np.int64)
121
+ }
122
+
123
+ # Run execution graph matrix calculations
124
+ intent_logits, ner_logits = self.session.run(None, onnx_feeds)
125
+
126
+ # Decode Intent array using Softmax calculation
127
+ logits_exp = np.exp(intent_logits[0])
128
+ probs = logits_exp / np.sum(logits_exp)
129
+ intent_idx = np.argmax(probs)
130
+
131
+ # Apply strict fallback thresholds
132
+ confidence = float(probs[intent_idx])
133
+ final_intent = self.intents[intent_idx] if confidence >= 0.65 else "unknown"
134
+
135
+ # Process and decode non-zero NER spans
136
+ ner_tags = np.argmax(ner_logits[0], axis=-1)
137
+ tokens = self.tokenizer.convert_ids_to_tokens(tokenized["input_ids"][0])
138
+
139
+ extracted_tokens = []
140
+ for token, tag in zip(tokens, ner_tags):
141
+ if token in [self.tokenizer.cls_token, self.tokenizer.sep_token, self.tokenizer.pad_token]:
142
+ continue
143
+ if tag > 0: # Valid B-DATA or I-DATA sequences
144
+ cleaned = token.replace("##", "")
145
+ if token.startswith("##") and extracted_tokens:
146
+ extracted_tokens[-1] += cleaned
147
+ else:
148
+ extracted_tokens.append(cleaned)
149
+
150
+ extracted_data = " ".join(extracted_tokens).strip()
151
+
152
+ return {
153
+ "intent": final_intent,
154
+ "confidence": confidence,
155
+ "extracted_data": extracted_data
156
+ }
157
+