File size: 9,471 Bytes
1540db2
 
 
 
 
 
 
 
 
 
 
 
 
3a81580
 
 
 
 
 
 
 
8516d4b
 
 
 
 
3a81580
 
 
8516d4b
3a81580
8516d4b
3a81580
 
 
 
 
 
8516d4b
3a81580
8516d4b
3a81580
8516d4b
3a81580
8516d4b
3a81580
 
 
 
 
 
 
 
 
 
 
 
 
 
8516d4b
3a81580
 
 
8516d4b
3a81580
 
8516d4b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3a81580
 
8516d4b
3a81580
8516d4b
 
3a81580
 
 
 
 
 
 
 
8516d4b
3a81580
 
 
 
8516d4b
3a81580
 
 
8516d4b
3a81580
 
8516d4b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3a81580
 
 
 
8516d4b
3a81580
 
 
 
8516d4b
3a81580
 
 
8516d4b
 
3a81580
 
 
 
8516d4b
3a81580
 
 
 
8516d4b
3a81580
8516d4b
 
3a81580
 
 
 
 
 
 
 
 
 
 
 
 
8516d4b
3a81580
 
 
 
8516d4b
3a81580
 
 
 
 
 
 
8516d4b
 
3a81580
 
 
8516d4b
3a81580
 
 
 
 
 
8516d4b
3a81580
 
 
8516d4b
3a81580
8516d4b
3a81580
 
 
8516d4b
 
 
 
 
 
3a81580
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
---
license: apache-2.0
base_model: indobenchmark/indobert-base-p1
language:
  - id
tags:
  - text-classification
  - pytorch
  - debate
  - fallacy-detection
pipeline_tag: text-classification
---


# Decida-small: 10-Pillar Automated Debate Adjudication Engine

[![License: MIT](https://img.shields.io/badge/License-MIT-blue.svg)](LICENSE)
[![Python 3.10+](https://img.shields.io/badge/Python-3.10%2B-brightgreen.svg)](https://www.python.org/)
[![Target Language](https://img.shields.io/badge/Language-Indonesian%20(Bahasa%20Indonesia)-red.svg)]()
[![Inference Latency](https://img.shields.io/badge/Latency-~4.2s%20(Local%20GPU)-orange.svg)]()
[![Zero External Daemons](https://img.shields.io/badge/Stack-100%25%20PyTorch%20Local-purple.svg)]()

**Decida-small** is an ultra-fast, 10-pillar automated debate adjudication engine tailored specifically for competitive Indonesian parliamentary and online debate tournaments.

Powered by fine-tuned **IndoBERT**, **Sentence-Transformers**, and local **DeBERTa NLI Cross-Encoders**, it evaluates multi-turn Affirmative vs. Opposition cases across 10 analytical dimensions, producing calibrated floating-point scores and constructive verbal feedback in approximately 4.2 seconds, without external LLM calls or cloud dependencies.

**Author: Christopher Alberto Hamonangan**

---

## Core Highlights

- **Sub-5s Local Inference:** Evaluates two teams across 6 dense multi-paragraph speeches in approximately 4.22 seconds on consumer GPU hardware.
- **Dedicated Indonesian NLP Stack:** Utilizes fine-tuned IndoBERT (`indobert-base-p1`) for fallacy classification and semantic evaluation tuned to formal Indonesian argumentation register.
- **Zero-Daemon / Self-Contained:** 100% Python and PyTorch. Eliminates external Ollama/HTTP dependencies, connection resets, and malformed JSON parsing timeouts.
- **Non-Judgmental Grounded Verbalizer:** Combinatorial slot-filling synthesis engine featuring 250+ modular stylistic variants that quotes real evidentiary claims from the winning team to deliver positive adjudication feedback.

---

## 10 Evaluation Pillars & Scoring Topology

The total final score is computed as a weighted sum with a 100% aggregate weight:

Final Score = Ξ£ (w_i Γ— P_i), for i = 1 to 10

| Pillar | Metric Dimension | Weight (\(w_i\)) | Underlying Method / Model |

|:---:|---|:---:|---|

| **P1** | **Kepadatan Semantik (Semantic Relevance)** | 10% | Dense cosine similarity vs. motion centroid via `all-MiniLM-L6-v2` |

| **P2** | **Integritas Logika (Fallacy & Logic)** | 15% | Fine-tuned IndoBERT Sequence Classifier (`./model_indobert_fallacy`) |

| **P3** | **Daya Respons/Bantahan (Rebuttal Clash)** | 10% | Batch local NLI cross-encoder (`contradiction` detection) |

| **P4** | **Struktur A-R-E-L (Argument Structure)** | 10% | Structural causality, assertion grounding, and empirical regex patterns |

| **P5** | **Pertarungan Dampak (Impact Weighing)** | 10% | Multi-layer trade-off and clash magnitude vector calculation |

| **P6** | **Alur Wacana / Dinamika (Discourse Drift)** | 5% | Trajectory velocity & consecutive paragraph transition matrices |

| **P7** | **Kepadatan Leksikal/Gaya (Lexical Register)** | 5% | Formal debate jargon index, lexical diversity (TTR), & structural depth |

| **P8** | **Konsistensi Internal Tim (Team Consistency)** | 10% | Pairwise premise cross-check via NLI + semantic drift penalties |

| **P9** | **Pertahanan & Komparasi (Resilience & Delta)** | 15% | 70% premise defense ratio + 30% comparative delta vs. opposition |

| **P10** | **Eksploitasi & Kepadatan (Targeting & Density)** | 10% | 60% core attack targeting ratio + 40% propositional info density |



---



## Benchmark Run Sample



```text

Mosi: "Dewan ini akan mewajibkan transisi 100% armada transportasi umum ke kendaraan listrik murni (EV) dalam waktu lima tahun ke depan."



Waktu Eksekusi: 4.22 detik



Metrik Pilar                  | TIM PRO | TIM KONTRA

----------------------------------------------------

P1: Kepadatan Semantik        |  82.65  |  80.92

P2: Integritas Logika         |  75.00  |  25.00

P3: Daya Respons/Bantahan     |  66.67  |  78.33

P4: Struktur A-R-E-L          |  52.50  |  59.17

P5: Pertarungan Dampak        |  57.85  |  83.38

P6: Alur Wacana / Dinamika    |  91.56  |  93.20

P7: Kepadatan Leksikal/Gaya   |  56.75  |  78.59

P8: Konsistensi Internal Tim  |  89.00  |  83.12

P9: Pertahanan & Komparasi    |  49.14  |  53.90

P10: Eksploitasi & Kepadatan  |  81.93  |  84.72

----------------------------------------------------

TOTAL SKOR AKHIR              |  69.10  |  67.39

====================================================



Pemenang: TIM AFIRMASI (PRO) (Margin: 1.71)



----------------------------------------------------

ULASAN ANALITIS DEWAN JURI (DECIDA-SMALL FEEDBACK):

----------------------------------------------------



Kemenangan dengan margin kompetitif 1.71 poin berhasil direngkuh oleh TIM AFIRMASI (PRO).



Penguasaan materi mendasar terbukti lewat susunan argumen yang utuh seperti terlihat saat menegaskan bahwa "Berdasarkan laporan Kementerian Lingkungan Hidup, lebih dari 65% polusi partikulat PM2.5 bersumber...", membangun alur berpikir yang meyakinkan.



Memasuki babak pembelaan dan penegasan posisi, alur penalaran tim mengalir secara organik (skor 89.00), di mana pembicara lanjutan bertindak sebagai akselerator kasus yang efektif.



Keunggulan paling menentukan hadir pada ketelitian berpikir tim, membentengi setiap kesimpulan dengan premis pendukung yang sah. Keberhasilan tim dalam menjaga beban pembuktian secara konsisten menjadi alasan fundamental di balik hasil ini.



====================================================

```



---



## Repository Structure



```text

decida-small/

β”‚

β”œβ”€β”€ pillar/

β”‚   β”œβ”€β”€ __init__.py

β”‚   β”œβ”€β”€ nli_client.py          # Local PyTorch batch NLI Cross-Encoder wrapper
β”‚   β”œβ”€β”€ verbalizer.py          # Combinatorial grounded feedback generator (~0.001s)
β”‚   β”œβ”€β”€ pillar1.py             # P1: Semantic density & motion relevance
β”‚   β”œβ”€β”€ pillar2.py             # P2: Logical integrity (IndoBERT fallacy detection)
β”‚   β”œβ”€β”€ pillar3.py             # P3: Rebuttal validity & direct clash
β”‚   β”œβ”€β”€ pillar4.py             # P4: A-R-E-L structure analysis
β”‚   β”œβ”€β”€ pillar5.py             # P5: Impact weighing & clash balance
β”‚   β”œβ”€β”€ pillar6.py             # P6: Discourse trajectory & flow
β”‚   β”œβ”€β”€ pillar7.py             # P7: Lexical density, register, & diversity
β”‚   β”œβ”€β”€ pillar8.py             # P8: Multi-turn internal team consistency
β”‚   β”œβ”€β”€ pillar9.py             # P9: Premise defense & comparative delta
β”‚   └── pillar10.py            # P10: Strategic premise exploitation & density
β”‚
β”œβ”€β”€ model_indobert_fallacy/    # Fine-tuned weights (or pull via Hugging Face Hub)
β”œβ”€β”€ model_minilm/              # Local sentence-transformer cache

β”‚

β”œβ”€β”€ .gitignoxre

β”œβ”€β”€ LICENSE

β”œβ”€β”€ README.md

β”œβ”€β”€ requirements.txt

└── run_judger.py              # CLI driver script with batch execution benchmark
```



---



## Installation & Setup



### 1. Prerequisites



- Python 3.10 or higher

- NVIDIA CUDA-compatible GPU (recommended for sub-5s performance) or modern multi-core CPU



### 2. Clone and Environment Setup



```bash

git clone https://github.com/USERNAME/decida-small.git

cd decida-small



python -m venv venv



# Windows (PowerShell)

.\venv\Scripts\Activate.ps1



# Linux / macOS

source venv/bin/activate



pip install --upgrade pip

pip install -r requirements.txt

```

### 3. Weights Configuration

Ensure the fine-tuned IndoBERT directory (`./model_indobert_fallacy`) and Sentence-Transformers directory (`./model_minilm`) exist locally, or update the model identifiers in `pillar/pillar2.py` and `run_judger.py` to point to your Hugging Face Model Hub repository:

```python

# pillar/pillar2.py



MODEL_PATH = "USERNAME_HF/decida-indobert-fallacy"

```

---

## Usage

Run the standard evaluation interface:

```bash

python run_judger.py

```

- **Mode 1 (Template):** Automatically loads the built-in EV fleet transition motion along with 3 Affirmative and 3 Opposition speeches for immediate profiling.
- **Mode 2 (Live):** Prompts for interactive speech-by-speech input with live multi-line terminal entry (`--end` to conclude debate turns).

---

## Verbalizer Architecture

Unlike LLM prompt generators that can hallucinate or produce demoralizing critiques, `pillar/verbalizer.py`:

1. Computes the maximum comparative margin across all 10 pillars.
2. Extracts verbatim evidentiary phrases directly from the winning team's text.
3. Assembles a 4-paragraph formal critique using a slot-filling combination engine with over 250 verified positive rhetorical options.
4. Completes execution in **less than 0.002 seconds**.

---

## Citation & License

This project is open-sourced under the [MIT License](LICENSE).

```bibtex

@software{decida_small_2026,

  author       = {Christopher Alberto Hamonangan},

  title        = {Decida-small: 10-Pillar Automated Debate Adjudication Engine for Indonesian Debate},

  year         = {2026},

  publisher    = {Huggingface},

  journal      = {Huggingface Organization},

  howpublished = {\url{https://huggingface.co/BhinnekaAI/Decida-small}

}

```