Nasim435/Qwen-3B-Automotive-20K

QWEN-3B MODEL CARD / AUTOMOTIVE DOMAIN ■ PRODUCTION READY QWEN-3B
AUTOMOTIVE BASE → Qwen/Qwen2.5-3B-Instruct  ·  METHOD → QLoRA / 4-bit NF4  ·  SAMPLES → 20,000 QLoRA Automotive MLflow LLM-as-a-Judge PEFT SFT Unsloth 4-bit NF4 bfloat16 TRL BitsAndBytes MIT License 01 OVERVIEW

Production-ready domain-adapted variant of Qwen2.5-3B-Instruct, fine-tuned on automotive instruction-following data using QLoRA with Unsloth optimization. Features comprehensive evaluation pipeline, dataset engineering, and experiment tracking for enterprise-grade LLM development.

Specialized toward automotive question answering, diagnostic explanations, vehicle maintenance assistance, and technical guidance. Trained on a curated subset of 20,000 samples from the BAAI automotive industry instruction dataset with advanced data quality controls.

Model Identity BASEQwen/Qwen2.5-3B-Instruct DEVELOPED BYAlibaba Cloud / Qwen Team FINE-TUNED BYNasim Raj Laskar LANGUAGEEnglish LICENSEMIT Capabilities ► Automotive Q&A — 96.6% data quality score ► Diagnostic Explanations & Troubleshooting ► Repair & Maintenance Guidance ► Vehicle Systems Knowledge ► Safety-focused Technical Instructions 02 TRAINING DATA & ENGINEERING

Fine-tuned on 20,000 curated samples from BAAI/IndustryInstruction_Automobiles with a comprehensive data engineering pipeline featuring automated quality controls and versioning.

Dataset Processing Pipeline FORMATQwen chat-template conversations STRUCTUREsystem → user → assistant SYSTEM PROMPT"You are an automotive expert assistant." SAMPLES~20,000 curated instruction pairs SPLITS90% train / 5% validation / 5% test Data Quality Metrics QUALITY SCORE96.6% average DUPLICATE RATE<0.01% AVG TOKENS114.6 per sample PROMPT TOKENS28.2 avg RESPONSE TOKENS86.4 avg ◆ DATA QUALITY FEATURES

Duplicate detection and removal (exact + near-duplicate)  ·  Quality scoring with flagging system  ·  Token length filtering (10–512 tokens)  ·  Malformed structure detection  ·  Automated quality reporting  ·  Dataset versioning with quality tracking

03 TRAINING CONFIGURATION MethodQLoRA Quantization4-bit NF4 Precisionbfloat16 OptimizerAdamW 8bit Max Seq Len512 LR ScheduleCosine ► LoRA Adapter Parameters Rank (r)16 Alpha (α)32 Dropout0.0 Trainable Params~30M Total Params~3.1B Infrastructure PLATFORMAWS SageMaker GPUNVIDIA L4 TRACKINGMLflow + DagsHub Libraries Transformers · TRL · PEFT · BitsAndBytes Accelerate · Datasets · Unsloth · MLflow ◆ TRAINING PIPELINE FEATURES

Validation-based early stopping  ·  Overfitting detection (threshold: 2.0)  ·  Best checkpoint selection via eval_loss  ·  Gradient checkpointing for memory efficiency  ·  MLflow experiment tracking with DagsHub  ·  Target modules: all attention + MLP projections

04 EVALUATION RESULTS

MULTI-METRIC EVALUATION SUITE · 20 SAMPLES · LLM-AS-A-JUDGE: GROQ LLAMA 3.3-70B

Core Performance PERPLEXITY5.47 BLEU SCORE16.3% SIMILARITY SCORE19.7% AVG LATENCY1,066 ms THROUGHPUT18.8 tok/sec LLM-as-a-Judge (0–10) HELPFULNESS7.4 / 10 CORRECTNESS8.3 / 10 COHERENCE8.8 / 10 INSTRUCTION FOLLOWING7.0 / 10 HALLUCINATION RISK ↓8.4 / 10 SAFETY9.2 / 10 GPU Metrics PEAK VRAM6.26 GB AVG GPU UTIL72.6% MAX GPU UTIL100.0% Training Throughput AVG TRAINING175 tok/sec MAX TRAINING643 tok/sec 05 PRODUCTION FEATURES Experiment Tracking & Monitoring ► MLflow integration with DagsHub remote tracking ► Automated parameter and metric logging ► GPU profiling with VRAM and power monitoring ► Runtime configuration capture ► Git metadata tracking Evaluation Framework ► Multi-metric evaluation suite ► LLM-as-a-Judge via external API (Groq) ► Pairwise comparison capabilities ► Automated post-training evaluation ► Performance benchmarking ◆ DATA ENGINEERING PIPELINE

Automated dataset quality analysis  ·  Duplicate detection and removal  ·  Quality scoring and filtering  ·  Dataset versioning system  ·  Comprehensive quality reporting

06 EXAMPLE USAGE Python

# Load model from Hugging Face Hub
from transformers import AutoTokenizer, AutoModelForCausalLM
model_name = "Nasim435/Qwen-3B-Automotive-20K"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model     = AutoModelForCausalLM.from_pretrained(
    model_name,
    device_map="auto",
    torch_dtype="auto"
)
prompt   = "Explain symptoms of a failing alternator and diagnostic steps."
messages = [{"role": "user", "content": prompt}]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)
inputs  = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(
    **inputs,
    max_new_tokens=200,
    temperature=0.7,
    top_p=0.9,
    do_sample=True,
    pad_token_id=tokenizer.eos_token_id
)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

07 LIMITATIONS & SAFETY ⚠ WARNING

• Research model — experimental fine-tune, not intended for production safety systems
• Hallucination risk — may generate inaccurate automotive advice (8.4/10 risk score)
• Safety critical — not suitable for safety-critical or professional mechanical decision-making
• Domain scope — trained on 20K samples; generalization beyond automotive may be limited
• Quality assurance — 96.6% data quality score with 263 flagged samples requiring review

08 TECHNICAL ARCHITECTURE Memory Optimization ► Unsloth FastLanguageModel integration ► 4-bit quantization with NF4 format ► Gradient checkpointing for memory efficiency ► Peak VRAM: 6.26 GB — consumer GPU compatible Performance Optimizations ► Fused attention kernels via Unsloth ► Optimized transformer implementations ► Efficient LoRA adapter injection ► 2–5x speedup over standard implementations ◆ MONITORING & OBSERVABILITY

Real-time GPU utilization tracking  ·  Memory usage profiling  ·  Training throughput monitoring  ·  Automated performance benchmarking

09 ACKNOWLEDGEMENTS Qwen TeamAlibaba Cloud BAAIDataset Contributors Hugging FaceEcosystem TRL / PEFTContributors UnslothContributors BitsAndBytesQuantization MLflow / DagsHubExperiment Tracking GroqLLM-as-a-Judge API NASIM RAJ LASKAR Qwen-3B-Automotive-20K
MIT License · English · 2026
Production-Ready Automotive AI Assistant