ArgParser-v3-Qwen1.5B

Version 3 of the ArgParser model series — continual fine-tuning with an additional argument-mining corpus to investigate the limits of domain transfer.

ArgParser-v3-Qwen1.5B extends ArgParser-v2 through continual training after introducing an additional academic argument-mining corpus (AAEC). The objective was to determine whether increasing the diversity of gold-standard argument annotations would improve generalization to real-world fact-checking data.

While this version achieved modest improvements on in-domain benchmarks, experiments demonstrated that academic argument-mining datasets alone were insufficient for effective transfer to PolitiFact-style claims. This finding directly motivated the development of ArgParser-v4.

---

Model Overview

Property	Value
Model	ArgParser-v3-Qwen1.5B
Base Model	Qwen/Qwen2.5-1.5B-Instruct
Task	Argument Structure Extraction
Training Method	Continual LoRA Fine-tuning
LoRA Rank	16
Author	Jayesh Choudhari
License	Apache-2.0

---

Training Configuration

ArgParser-v3 continues training from Version 2 after incorporating an additional argument-mining dataset.

Training datasets:

• AbstRCT
• Microtext
• CDCP
• PERSPECTRUM
• AAEC (402 persuasive essays)

Training configuration:

• Training Samples: 1,823
• Continual Training: 1 additional epoch
• Total Effective Learning: ~4 epochs including Version 2
• LoRA Rank: 16
• Hardware: NVIDIA GTX 1080 Ti
• Training Time: ~5.5 hours

---

Performance

Average held-out Component F1:

Component F1: 0.229

Compared with Version 2:

• Small overall improvement (0.219 → 0.229)
• Improved performance on:
  • Microtext
  • AbstRCT
• Slight regression on:
  • PERSPECTRUM (0.056 → 0.034)

These results suggest diminishing returns when simply increasing the amount of academic argument-mining data.

---

Cross-Domain Evaluation

A key experiment evaluated ArgParser-v3 on the target LIARArg dataset containing PolitiFact-style political claims.

Results showed:

• Approximately 83% empty predictions on the initial evaluation subset.
• Remaining outputs were often incomplete or fragmented.
• The experiment was terminated early after confirming poor cross-domain transfer.

This demonstrated that academic argument-mining corpora alone do not adequately prepare small language models for real-world political fact-checking tasks.

These findings motivated the development of ArgParser-v4, which introduces target-domain silver annotations generated by a larger teacher model.

---

Intended Task

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

• Claims
• Premises
• Citations
• Supporting relations
• Attacking relations

Outputs follow a structured JSON format suitable for downstream argument mining pipelines.

---

Usage

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

base_id = "Qwen/Qwen2.5-1.5B-Instruct"
adapter_id = "iamjayeshc/ArgParser-v3-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:

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-v3 is the third model in the ArgParser development series.

Other versions:

• ArgParser-v1-Qwen0.5B

  • Initial full fine-tuning baseline
  • https://huggingface.co/iamjayeshc/ArgParser-v1-Qwen0.5B

• ArgParser-v2-Qwen1.5B

  • Larger base model with LoRA fine-tuning
  • https://huggingface.co/iamjayeshc/ArgParser-v2-Qwen1.5B

• ArgParser-v4-Qwen1.5B (Recommended)

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

ArgParser-v4 incorporates silver annotations generated by a large teacher model and substantially improves transfer to PolitiFact-style argument extraction.

---

Limitations

This version demonstrates that expanding academic training data alone is insufficient for robust cross-domain argument extraction.

It may struggle with:

• Political claims
• Informal social media language
• Long-form real-world arguments

The model is primarily released for research reproducibility and comparison with subsequent versions.

---

Acknowledgements

This model is part of the ArgParser research project investigating lightweight argument-structure extraction through continual training, domain adaptation, and knowledge distillation.

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 is planned.