ArgParser-v3

v2's adapter continued for one more epoch after adding a fifth corpus: AAEC (402 persuasive essays, ~6000 argument components). ~5.5 hours on the same GTX 1080 Ti.

Held-out component-F1: 0.229, a marginal improvement over v2's 0.219. Microtext and AbstRCT nudged up; PERSPECTRUM slightly regressed (0.056 → 0.034). Adding more of the same kind of extractive academic gold hits diminishing returns pretty quickly.

I also tried v3 on the actual LIARArg parse — the whole point of the project — and hit an 83% empty rate on the first 64 rows. Real outputs were fragmentary ("is not clear" as a claim). Killed the run after that; it was obvious this variant couldn't do cross-domain transfer to Politifact-style claims. The five academic argument-mining corpora aren't enough on their own to bridge that gap.

That result motivated v4 — adding silver labels from a large teacher (gpt-oss-120b) on 2,123 LIARArg training articles, with Chain-of-Thought reasoning traces preserved through training. v4 gets Phase 1 integration F1 = 0.217, closes 33% of the gold-parser gap.

For actual use, go to v4. This one exists for the ablation record.

Config

• Base: Qwen/Qwen2.5-1.5B-Instruct
• Method: LoRA r=16, continual from v2
• Data: 5 gold corpora (AAEC added), 1,823 records
• Epochs: 1 continual (~4 epochs of learning total including v2)
• Wall clock: 5.5 h

Usage

from peft import PeftModel
from transformers import AutoModelForCausalLM
base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-1.5B-Instruct")
model = PeftModel.from_pretrained(base, "properexit/ArgParser-v3")

License

Apache 2.0.