vidavox/Qwen3-SKK-32B-SFT-LoRA

LoRA adapter for a Qwen3-32B base model, fine-tuned on SKK’s KSMI document data for domain-specific question answering.

This repository contains only the LoRA adapter weights. To use the model, you must load a compatible Qwen3-32B base model and then attach this adapter with PeftModel.from_pretrained.

---

Model Details

• Base model (expected): Qwen/Qwen3-32B (or another compatible Qwen3-32B variant)
• Adapter type: LoRA via PEFT
• Task: Supervised fine-tuning for assistant-style answers grounded in SKK’s KSMI document data.
• Languages: Primarily Bahasa Indonesia and English in a technical / regulatory context.

This adapter is intended to be used in SKK’s internal systems for answering questions based on KSMI and related upstream oil & gas regulatory documents.

---

Usage (PEFT / LoRA)

⚠️ This repo does not include the base model. You must:

1. Load the Qwen3-32B base model.
2. Wrap it with PeftModel.from_pretrained using this adapter.

1. Install dependencies

pip install "transformers>=4.43.0" peft accelerate bitsandbytes

2. Load the base Qwen3-32B model

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

BASE_MODEL_ID = "Qwen/Qwen3-32B"  # adjust if you used a different Qwen3-32B ID

tokenizer = AutoTokenizer.from_pretrained(
    BASE_MODEL_ID,
    trust_remote_code=True,
)

base_model = AutoModelForCausalLM.from_pretrained(
    BASE_MODEL_ID,
    device_map="auto",
    torch_dtype=torch.bfloat16,   # or "auto"
    trust_remote_code=True,
)

3. Attach the LoRA adapter with PeftModel

from peft import PeftModel

ADAPTER_ID = "vidavox/Qwen3-SKK-32B-SFT-LoRA"

model = PeftModel.from_pretrained(
    base_model,
    ADAPTER_ID,
    torch_dtype=torch.bfloat16,   # should match the base model dtype
)
model.eval()

PeftModel.from_pretrained loads the adapter configuration and weights and applies them to the provided base model.

4. Chat-style inference (Qwen3 chat template)

Qwen models use a chat template accessed via tokenizer.apply_chat_template.

messages = [
    {
        "role": "system",
        "content": "You are an assistant specialized in SKK KSMI documents.",
    },
    {
        "role": "user",
        "content": "Jelaskan secara ringkas tahapan proses persetujuan POD berdasarkan KSMI.",
    },
]

inputs = tokenizer.apply_chat_template(
    messages,
    tokenize = False,
    add_generation_prompt=True,
    enable_thinking = False, # Disable thinking
)

model_inputs = tokenizer([input_text], return_tensors="pt").to(model.device)
generated_ids = model.generate(
    **model_inputs,
    max_new_tokens=max_new_tokens,
    do_sample=False,
    temperature=0.1,
    top_p=0.95,
)
output_ids = generated_ids[0][len(model_inputs.input_ids[0]):].tolist() 
response = tokenizer.decode(output_ids, skip_special_tokens=True)
print(response)

5. Optional: 4-bit loading for constrained VRAM

To run base + adapter on a single 24 GB GPU (e.g. RTX 3090), you can use 4-bit quantization with bitsandbytes:

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

BASE_MODEL_ID = "Qwen/Qwen3-32B"
ADAPTER_ID = "vidavox/Qwen3-SKK-32B-SFT-LoRA"

bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_compute_dtype=torch.bfloat16,
    bnb_4bit_use_double_quant=True,
    bnb_4bit_quant_type="nf4",
)

tokenizer = AutoTokenizer.from_pretrained(
    BASE_MODEL_ID,
    trust_remote_code=True,
)

base_model = AutoModelForCausalLM.from_pretrained(
    BASE_MODEL_ID,
    device_map="auto",
    quantization_config=bnb_config,
    trust_remote_code=True,
)

model = PeftModel.from_pretrained(base_model, ADAPTER_ID)
model.eval()

The generation code is then the same as in step 4.

---

Training Data

The adapter was trained on SKK’s internal KSMI document data, formatted as instruction-style examples.

• Train split: 2223 examples
• Validation split: 247 examples
• Test split: 50 examples

The data consists of questions and instructions grounded in KSMI and related SKK upstream oil & gas regulations, with answers written to be faithful to the underlying documents. The dataset is private and not released with this model.

---

Evaluation (SDA on test data)

Evaluation was performed on the 50-sample test set using an SDA-style pipeline that combines automatic metrics (text/semantic similarity) and human-oriented quality scores.

Metric summary

Metric	Mean (test)	Scale / note
BERTScore F1	0.845	0–1, higher = better semantic similarity
Correctness	5.96	1–10, higher = logically correct answers
Completeness	5.10	1–10, higher = more required information covered
Factuality	7.08	1–10, higher = fewer factual errors
Structure	7.82	1–10, higher = better organization / formatting
Hallucination resistance	7.08	1–10, higher = less hallucination

These metrics are computed on a small, domain-specific test set and should be interpreted as indicative of quality on KSMI-style questions, not as general-purpose benchmarks.

---

Intended Use (High-Level)

• Primary use: Internal SKK assistant systems answering questions grounded in KSMI and related upstream O&G regulations.
• Not intended for: General-purpose open-domain chat, safety-critical decision making, or use outside the domain without additional evaluation and alignment.