DuckerMaster's picture
docs: release model card, benchmark figures, and architecture metadata
d640b95 verified
|
Raw History Blame Contribute Delete
6.42 kB
---
language:
- th
license: apache-2.0
tags:
- token-classification
- ner
- temporal-extraction
- wangchanberta
- roberta
- thai-nlp
base_model: airesearch/wangchanberta-base-att-spm-uncased
datasets:
- DuckerMaster/Thai-Natural-DateTime-Dataset
pipeline_tag: token-classification
metrics:
- f1
- precision
- recall
- accuracy
model-index:
- name: DuckerTime-Normalize-Thai
results:
- task:
type: token-classification
name: Named Entity Recognition
dataset:
name: Thai Natural DateTime Dataset (v1.1)
type: DuckerMaster/Thai-Natural-DateTime-Dataset
metrics:
- name: In-Distribution F1
type: f1
value: 1.0000
- name: Out-of-Distribution F1
type: f1
value: 0.9692
- name: Out-of-Distribution Recall
type: recall
value: 0.9957
- name: Out-of-Distribution Precision
type: precision
value: 0.9441
---
# DuckerTime-Normalize-Thai (WangchanBERTa Temporal Entity Slicer)
Fine-tuned `wangchanberta-base-att-spm-uncased` for token-level temporal boundary extraction (`B-TIME`, `I-TIME`, `O`) across colloquial Thai sentences, conversational chat slang, and compound date-time structures.
Paired Dataset: [DuckerMaster/Thai-Natural-DateTime-Dataset](https://huggingface.co/datasets/DuckerMaster/Thai-Natural-DateTime-Dataset)
Designed as the neural frontend for the End-to-End Hybrid Architecture:
1. **WangchanBERTa Frontend**: Extracts precise character and token spans from raw unstructured text.
2. **Deterministic Rule Engine Backend**: Parses the extracted temporal span into valid ISO-8601 points and ranges in sub-millisecond execution.
## Model Performance & Empirical Benchmark
![Empirical Benchmark Comparison](live_t4_benchmark_comparison.png)
Evaluated against disjoint stratified benchmarks and tested side-by-side with rule baselines and instruction-tuned LLMs on NVIDIA T4 hardware.
### Benchmark Summary (Batch Size = 1, Streaming Latency)
| System / Model Architecture | Parameter Size | Hardware Execution | OOD Accuracy / F1 | Mean Latency (ms) | Production SLA (<50ms) |
| :--- | :---: | :---: | :---: | :---: | :---: |
| **dateparser** (Heuristic Multilingual) | 0 | CPU | 6.0% | 12.4 ms | Pass |
| **PyThaiNLP** (Rule Parser) | 0 | CPU | 18.7% | 1.8 ms | Pass |
| **Qwen2.5-3B-Instruct** (4-bit NF4) | 3B | GPU (T4) | 3.3% | 383.3 ms | Fail (Over 7x threshold) |
| **Typhoon 30B (Cloud REST API)** | 30B | Cloud Server | 47.3% | 2,753.9 ms | Fail (Over 55x threshold) |
| **Our Hybrid System (WangchanBERTa + Rule)** | **106M** | **GPU (T4)** | **96.9%** | **10.2 ms** | **Pass (Optimal Frontier)** |
### Architectural Insights
- **Why Pure LLMs Fail on Strict Temporal Extraction**: Generative LLMs are prone to conversational framing (e.g. outputting conversational preambles or rephrasing targets) rather than returning strict token spans. This results in high parsing failure rates (3.3% exact match) and high latency (383ms+ per request).
- **Why Hybrid Architecture Wins**: Token classification via WangchanBERTa isolates target temporal spans with 99.57% Recall on out-of-distribution text. Passing this span to a deterministic rule engine guarantees valid ISO-8601 timestamps without hallucination.
### Accuracy vs. Latency Pareto Frontier
![Accuracy vs Latency Frontier](accuracy_vs_latency_frontier.png)
## Model Specifications and Hyperparameters
- **Base Architecture**: RoBERTa / CamemBERT architecture pre-trained on Thai Wikipedia, news, and social text (`airesearch/wangchanberta-base-att-spm-uncased`).
- **Target Entities**:
- `O`: Outside temporal entity.
- `B-TIME`: Beginning of temporal expression.
- `I-TIME`: Continuation of temporal expression.
- **Training Epochs**: 4
- **Batch Size**: 16 (per device)
- **Optimizer**: AdamW (Learning Rate: 3e-5, Weight Decay: 0.01)
- **Warmup Steps**: 50
- **FP16 Mixed Precision**: Enabled
- **Inference Robustness**: Multi-token stitching with confidence thresholding (0.35) to handle spaced particles and chat modifiers.
## Intended Use and Pipeline Example
```python
import torch
from transformers import AutoTokenizer, AutoModelForTokenClassification
model_name = "DuckerMaster/DuckerTime-Normalize-Thai"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForTokenClassification.from_pretrained(model_name)
text = "ขอเลื่อนเป็นวันพุธหน้าช่วงบ่ายแก่ๆ สักสี่โมงเย็นได้ไหมคะ"
inputs = tokenizer(text, return_tensors="pt")
with torch.no_grad():
logits = model(**inputs).logits
preds = torch.argmax(logits, dim=2)[0].tolist()
tokens = tokenizer.convert_ids_to_tokens(inputs["input_ids"][0])
extracted_tokens = [tok for tok, pred in zip(tokens, preds) if pred != 0]
print("Extracted Temporal Tokens:", extracted_tokens)
```
## Limitations and Negative Sample Handling
- The model is fine-tuned to recognize temporal expressions in Thai. It actively suppresses non-temporal numeric entities (monetary sums, telephone sequences, tracking numbers) due to negative sampling in training data.
- For non-Thai text, fallback to multilingual parsers is recommended.
- Pair with `thai-temporal-normalizer` core package for full ISO-8601 normalization.
## Citation
If you use this model in your research or production systems, please cite:
```bibtex
@software{duckertime_thai_2026,
author = {DuckerMaster},
title = {DuckerTime-Normalize-Thai: Fine-Tuned WangchanBERTa for Colloquial Thai Temporal NLU},
year = {2026},
publisher = {HuggingFace},
howpublished = {\url{https://huggingface.co/DuckerMaster/DuckerTime-Normalize-Thai}}
}
```
## References
- **WangchanBERTa**: Lowphansirikul, L., et al. (2021). WangchanBERTa: Pretrained language models for Thai language. *arXiv preprint arXiv:2101.09635*.
- **ISO-TimeML**: Pustejovsky, J., et al. (2010). ISO-TimeML: An International Standard for Semantic Annotation. *ISO/TC 37/SC 4*.
- **SUTime**: Chang, A. X., & Manning, C. D. (2012). SUTime: A library for recognizing and normalizing time expressions. In *LREC 2012*.
- **HeidelTime**: Strötgen, J., & Gertz, M. (2010). HeidelTime: High quality rule-based extraction and normalization of temporal expressions. In *SemEval 2010*.