Update README

README.md

@@ -2,23 +2,45 @@
 license: apache-2.0
 base_model: Qwen/Qwen2.5-1.5B-Instruct
 library_name: peft
-tags:
-  - argument-mining
-  - fact-checking
-  - lora
-  - qwen
+tags: [argument-mining, fact-checking, lora, qwen]
 language: [en]
 pipeline_tag: text-generation
 ---
 
-# ArgRAG argrag-phase2beta-v2
+# ArgParser-v2
 
-**Placeholder — full model card coming.**
+Same training data as v1 (four argument-mining corpora, 1,494 records)
+but a larger base and LoRA instead of full fine-tune. Qwen-1.5B with
+LoRA r=16 (α=32, dropout 0.05, target `q_proj,k_proj,v_proj,o_proj`).
+3 epochs, about 13.5 hours on a GTX 1080 Ti.
 
-v2: LoRA r=16 on 4 gold corpora (AAEC/AbstRCT/CDCP/PERSPECTRUM/Microtext), fresh
+Held-out component-F1: **0.219** — roughly double v1. Microtext premise
+F1 went from 0.000 to 0.680. AbstRCT empty rate 75% → 50%. Scale plus
+LoRA plus longer training context are the dominant levers, and this
+was the run that made that obvious.
+
+Not the best model in the series. For actual use pick
+[ArgParser-v4](https://huggingface.co/properexit/ArgParser-v4) instead — it adds cross-domain transfer to
+LIARArg-style political claims, which is what most people probably
+care about.
+
+## Config
 
 - Base: `Qwen/Qwen2.5-1.5B-Instruct`
-- Method: LoRA r=16, alpha=32, target `q_proj,k_proj,v_proj,o_proj`
-- Repository: (link to be added)
+- Method: LoRA r=16 (4.4M trainable params)
+- Data: 4 gold corpora, 1,494 records
+- Epochs: 3
+- Wall clock: 13.5 h
+
+## Usage
+
+```python
+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-v2")
+```
+
+## License
 
-Detailed evaluation, usage, and limitations to be added.
+Apache 2.0.