Update README.md

README.md

@@ -1,7 +1,10 @@
 ---
 base_model: unsloth/qwen2.5-1.5b-instruct-unsloth-bnb-4bit
 library_name: peft
-model_name: sft_warmup
+language:
+- en
+license: apache-2.0
+pipeline_tag: text-generation
 tags:
 - base_model:adapter:unsloth/qwen2.5-1.5b-instruct-unsloth-bnb-4bit
 - lora
@@ -9,30 +12,59 @@ tags:
 - transformers
 - trl
 - unsloth
-licence: license
-pipeline_tag: text-generation
+- openenv
+- adversarial-robustness
+- structured-extraction
+- json-schema
 ---
 
-# Model Card for sft_warmup
+# Extractor (SFT warmup) — Adversarial Structured-Extraction Arena
+
+This model repo hosts the **SFT warmup LoRA adapter** trained for the OpenEnv project **Adversarial Structured-Extraction Arena**: an adversary perturbs messy documents/schemas (under a budget) and the extractor must output **valid JSON** matching a target schema.
 
-This model is a fine-tuned version of [unsloth/qwen2.5-1.5b-instruct-unsloth-bnb-4bit](https://huggingface.co/unsloth/qwen2.5-1.5b-instruct-unsloth-bnb-4bit).
-It has been trained using [TRL](https://github.com/huggingface/trl).
+## Links (submission)
 
-## Quick start
+- **GitHub repo**: https://github.com/Hardikjha09/openenv-adversarial-extraction-arena
+- **Runnable Space**: https://huggingface.co/spaces/HardikJha/extraction-arena
+- **Colab (re-run training)**: https://colab.research.google.com/github/Hardikjha09/openenv-adversarial-extraction-arena/blob/main/notebooks/Train_Extractor_Colab.ipynb
+- 
+## Evidence (plots + logs)
+
+- **Training loss**: https://huggingface.co/HardikJha/extractor-aea/blob/main/plots/sft_loss.png
+- **Eval reward (moving average)**: https://huggingface.co/HardikJha/extractor-aea/blob/main/plots/rewards.png
+- **Eval Elo**: https://huggingface.co/HardikJha/extractor-aea/blob/main/plots/elo_ratings.png
+- **Eval metrics JSON**: https://huggingface.co/HardikJha/extractor-aea/blob/main/eval_metrics.json
+- **SFT trainer log (raw JSON)**: https://huggingface.co/HardikJha/extractor-aea/blob/main/trainer_log_history.json
+- 
+## What this checkpoint is
+
+- **Base model**: `unsloth/Qwen2.5-1.5B-Instruct` (4-bit Unsloth bundle: `unsloth/qwen2.5-1.5b-instruct-unsloth-bnb-4bit`)
+- **Adapter**: LoRA (`peft`), saved from `training/sft_warmup.py`
+- 
+## Quick start (load base + adapter)
 
 ```python
-from transformers import pipeline
+import torch
+from transformers import AutoModelForCausalLM, AutoTokenizer
+from peft import PeftModel
 
-question = "If you had a time machine, but could only go to the past or the future once and never return, which would you choose and why?"
-generator = pipeline("text-generation", model="None", device="cuda")
-output = generator([{"role": "user", "content": question}], max_new_tokens=128, return_full_text=False)[0]
-print(output["generated_text"])
-```
+base_model_id = "unsloth/Qwen2.5-1.5B-Instruct"
+adapter_id = "HardikJha/extractor-aea"
 
-## Training procedure
+tokenizer = AutoTokenizer.from_pretrained(adapter_id, trust_remote_code=True)
 
- 
+base = AutoModelForCausalLM.from_pretrained(
+    base_model_id,
+    torch_dtype=torch.float16,
+    device_map="auto",
+    trust_remote_code=True,
+)
+model = PeftModel.from_pretrained(base, adapter_id)
+```
 
+## Training procedure
+- **Objective**:  supervised JSON extraction formatting aligned to the repo’s extractor prompt (`training/prompts.py`)
+- **Framework**: TRL SFTTrainer + Unsloth FastLanguageModel (see `training/sft_warmup.py`)
 
 This model was trained with SFT.
 
@@ -47,8 +79,6 @@ This model was trained with SFT.
 
 ## Citations
 
-
-
 Cite TRL as:
     
 ```bibtex