v7: r=96, alpha=192, lr=1e-6, max_len=1024, epoch=2 + DPO (README written by author)

README.md

@@ -2,41 +2,44 @@
 base_model: Qwen/Qwen3-4B-Instruct-2507
 datasets:
 - u-10bei/structured_data_with_cot_dataset_512_v2
+- u-10bei/dpo-dataset-qwen-cot
 language:
 - en
 license: apache-2.0
 library_name: peft
 pipeline_tag: text-generation
 tags:
-- qlora
-- lora
-- structured-output
+- dpo
+- unsloth
+- qwen
+- alignment
 ---
 
-#qwen3-4b-structured-output-lora
+# qwen3-4b-structured-sft-dpo
 
-This repository provides a **LoRA adapter** fine-tuned from
-**Qwen/Qwen3-4B-Instruct-2507** using **QLoRA (4-bit, Unsloth)**.
+This model is fine-tuned from **Qwen/Qwen3-4B-Instruct-2507** using SFT, then further optimized with **Direct Preference Optimization (DPO)** via the **Unsloth** library.
 
-This repository contains **LoRA adapter weights only**.
-The base model must be loaded separately.
+This repository contains **LoRA adapter weights**. The base model must be loaded separately with PeftModel.
 
 ## Training Objective
-
-This adapter is trained to improve **structured output accuracy**
-(JSON / YAML / XML / TOML / CSV).
-
-Loss is applied only to the final assistant output,
-while intermediate reasoning (Chain-of-Thought) is masked.
+This model has been optimized using DPO to better align responses with preferred outputs, with a primary focus on structured output quality and format adherence based on the preference dataset.
 
 ## Training Configuration
+**Stage 1 - SFT:**
+- **Dataset**: u-10bei/structured_data_with_cot_dataset_512_v2
+- **Method**: QLoRA (4-bit)
+- **Max sequence length**: 1024
+- **Epochs**: 2
+- **Learning rate**: 1e-06
+- **LoRA**: r=96, alpha=192
 
-- Base model: Qwen/Qwen3-4B-Instruct-2507
-- Method: QLoRA (4-bit)
-- Max sequence length: 1024
-- Epochs: 2
-- Learning rate: 1e-06
-- LoRA: r=96, alpha=192
+**Stage 2 - DPO:**
+- **Dataset**: u-10bei/dpo-dataset-qwen-cot
+- **Method**: DPO (Direct Preference Optimization)
+- **Epochs**: 1
+- **Learning rate**: 1e-07
+- **Beta**: 0.1
+- **Max sequence length**: 1024
 
 ## Usage
 
@@ -52,14 +55,14 @@ tokenizer = AutoTokenizer.from_pretrained(base)
 model = AutoModelForCausalLM.from_pretrained(
     base,
     torch_dtype=torch.float16,
-    device_map="auto",
+    device_map="auto"
 )
 model = PeftModel.from_pretrained(model, adapter)
-```
 
-## Sources & Terms (IMPORTANT)
+```
 
-Training data: u-10bei/structured_data_with_cot_dataset_512_v2
+## Sources & License (IMPORTANT)
 
+* **Training Data**: [u-10bei/structured_data_with_cot_dataset_512_v2], [u-10bei/dpo-dataset-qwen-cot]
 Dataset License: MIT License. This dataset is used and distributed under the terms of the MIT License.
 Compliance: Users must comply with the MIT license (including copyright notice) and the base model's original terms of use.

adapter_config.json

@@ -33,13 +33,13 @@
   "rank_pattern": {},
   "revision": null,
   "target_modules": [
-    "o_proj",
-    "gate_proj",
     "up_proj",
     "q_proj",
     "k_proj",
+    "v_proj",
+    "o_proj",
     "down_proj",
-    "v_proj"
+    "gate_proj"
   ],
   "target_parameters": null,
   "task_type": "CAUSAL_LM",

adapter_model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:d6fb93374d50edd5a452ddc507d0dc809bc7fffef75ce0827d1059a7ffc87906
+oid sha256:6a9e39256bfde27eff3909f5d8c9d4fb46df82bf341e5ec0e2bd37ac7bd5a47e
 size 792791480

tokenizer_config.json

@@ -233,7 +233,7 @@
   "extra_special_tokens": {},
   "model_max_length": 262144,
   "pad_token": "<|vision_pad|>",
-  "padding_side": "right",
+  "padding_side": "left",
   "split_special_tokens": false,
   "tokenizer_class": "Qwen2Tokenizer",
   "unk_token": null