README.md

---
license: apache-2.0
base_model: Qwen/Qwen2.5-1.5B-Instruct
library_name: peft
language:
- en
pipeline_tag: text-generation
tags:
- argument-mining
- fact-checking
- information-extraction
- lora
- qwen2
- peft
- research
---

# ArgParser-v2-Qwen1.5B

> **Version 2 of the ArgParser model series** — introducing LoRA fine-tuning and a larger Qwen model, resulting in a substantial improvement over the initial baseline.

ArgParser-v2-Qwen1.5B builds upon the foundation established by **ArgParser-v1** by replacing the 0.5B base model with **Qwen2.5-1.5B-Instruct** and adopting **LoRA (Low-Rank Adaptation)** for parameter-efficient fine-tuning.

While trained on the same four academic argument-mining datasets, this version demonstrates that scaling the base model and using LoRA significantly improves structured argument extraction performance.

---

# Model Overview

| Property | Value |
|----------|-------|
| Model | ArgParser-v2-Qwen1.5B |
| Base Model | Qwen/Qwen2.5-1.5B-Instruct |
| Task | Argument Structure Extraction |
| Training Method | LoRA Fine-tuning |
| LoRA Rank | 16 |
| Author | **Jayesh Choudhari** |
| License | Apache-2.0 |

---

# Training Configuration

Training was performed using the same four manually annotated argument-mining datasets used in Version 1:

- AbstRCT
- Microtext
- CDCP
- PERSPECTRUM

Training configuration:

- **Training Samples:** 1,494
- **Epochs:** 3
- **LoRA Rank:** 16
- **LoRA Alpha:** 32
- **Dropout:** 0.05
- **Target Modules:** `q_proj`, `k_proj`, `v_proj`, `o_proj`
- **Trainable Parameters:** ~4.4 Million
- **Hardware:** NVIDIA GTX 1080 Ti
- **Training Time:** ~13.5 hours

---

# Performance

Average held-out Component F1:

**Component F1:** **0.219**

Compared with **ArgParser-v1**, this represents approximately a **2× improvement**.

Notable improvements include:

- Microtext Premise F1:
  - **0.000 → 0.680**

- AbstRCT Empty Prediction Rate:
  - **75% → 50%**

These results demonstrate the benefits of combining a larger base model with parameter-efficient fine-tuning.

---

# Intended Task

The model extracts structured argumentative information from input text, including:

- Claims
- Premises
- Citations
- Supporting relations
- Attacking relations

The output is designed for downstream argument mining and fact-checking pipelines.

---

# Usage

```python
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

base_id = "Qwen/Qwen2.5-1.5B-Instruct"
adapter_id = "iamjayeshc/ArgParser-v2-Qwen1.5B"

tokenizer = AutoTokenizer.from_pretrained(base_id)

base = AutoModelForCausalLM.from_pretrained(
    base_id,
    torch_dtype=torch.float16,
    device_map="auto"
)

model = PeftModel.from_pretrained(base, adapter_id)
```

Example prompt:

```python
instruction = (
    "Extract all argument components and relations from the text. "
    "Return strict JSON containing "
    "claim_components, premise_components, citation_components and relations."
)

text = "The Obama administration is putting Border Patrol agents in a chokehold."
```

---

# Model Series

ArgParser-v2 is the second model in the ArgParser development series.

Other versions:

- **ArgParser-v1-Qwen0.5B**
  - Baseline full fine-tuning
  - https://huggingface.co/iamjayeshc/ArgParser-v1-Qwen0.5B

- **ArgParser-v3-Qwen1.5B**
  - Continual training with an additional argument-mining corpus
  - https://huggingface.co/iamjayeshc/ArgParser-v3-Qwen1.5B

- **ArgParser-v4-Qwen1.5B (Recommended)**
  - Final distilled model with target-domain adaptation
  - https://huggingface.co/iamjayeshc/ArgParser-v4-Qwen1.5B

For most applications, **ArgParser-v4** is recommended, as it incorporates cross-domain adaptation for PolitiFact/LIARArg-style claims.

---

# Limitations

Although significantly stronger than Version 1, this model still relies exclusively on academic argument-mining corpora for training.

Consequently, it may struggle with:

- Informal social media language
- Political claims
- Domain-specific argument structures

These limitations motivated the development of Versions 3 and 4.

---

# Acknowledgements

This model is part of the **ArgParser** research project investigating efficient argument-structure extraction using progressively improved lightweight language models.

Released by **Jayesh Choudhari**.

---

# Citation

If you use this model in your research, please cite the associated project when available.

A formal technical report / preprint will be released in the future.