jenosize-qwen3-4b

Model Card for Jenosize Qwen3 4B (Style Transfer)

This model is fine-tuned to generate creative business trend analysis and marketing articles in the professional and creative tone of Jenosize Digital Agency.

Model Details

Model Description

This model is a fine-tuned version of the iapp/chinda-qwen3-4b base model, adapted using Low-Rank Adaptation (LoRA) via the Unsloth library. It was specifically trained on bilingual (Thai and English) articles scraped from the Jenosize Ideas blog. The primary goal of this model is to perform Style Transfer, adopting the specific brand voice, vocabulary, and analytical style of Jenosize.

• Developed by: mix8645 (Worachot Chanmueang)
• Model type: Causal Language Model (Fine-tuned with LoRA)
• Language(s) (NLP): Thai (th), English (en)
• License: Apache 2.0 (Inherited from base model)
• Finetuned from model: iapp/chinda-qwen3-4b

Model Sources

• Repository: https://huggingface.co/mix8645/jenosize-qwen3-4b

Uses

Direct Use

The model is intended to be used by digital marketers, content creators, and developers to generate drafts for:

• Trend analysis articles.
• Business strategy and marketing insights.
• Phygital and AI technology trends in an agency tone.

Out-of-Scope Use

• Serving as a factual knowledge base (the model is optimized for "style" and may hallucinate specific facts or dates).
• Generating harmful, biased, or explicit content.

Bias, Risks, and Limitations

While the model captures the professional tone of Jenosize, it is still an AI and its outputs should be reviewed by a human before publication. It relies on the training data's perspective (digital agency viewpoint) and may over-index on specific marketing buzzwords.

How to Get Started with the Model

Use the code below to get started with the model using the Hugging Face requests API or transformers.

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "mix8645/jenosize-qwen3-4b"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name)

instruction = "เขียนบทความวิเคราะห์เทรนด์ธุรกิจในสไตล์ Jenosize ที่มีความคิดสร้างสรรค์และเป็นมืออาชีพ"
topic = "อนาคตของการทำการตลาดด้วย AI ในปี 2026"

prompt = f"### Instruction:\n{instruction}\n\n### Input:\nTopic: {topic}\n\n### Response:\n"
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=800)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Training Details

Training Data

The dataset consists of business and technology articles scraped directly from the Jenosize Ideas website. The data was augmented using chunking techniques and formatted into the Alpaca instruction format, categorizing inputs by Thai and English prompts.

Training Procedure

The model was trained using the unsloth library for faster and memory-efficient fine-tuning on a Google Colab T4 GPU.

Training Hyperparameters

• Training regime: 4-bit quantization (QLoRA)
• LoRA Rank (r): 16
• LoRA Alpha: 16
• Target Modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
• Learning Rate: 3e-5
• LR Scheduler: Cosine
• Epochs: 8
• Optimizer: adamw_8bit
• Max Sequence Length: 2048
• Batch Size: 2 (with Gradient Accumulation Steps = 4)

Environmental Impact

• Hardware Type: 1x NVIDIA T4 GPU (Google Colab)
• Hours used: ~0.5 hours
• Compute Region: Global

Technical Specifications

Compute Infrastructure

Trained on Google Colab (Free Tier).

Software

• unsloth
• transformers
• peft
• trl

Model Card Authors

Worachot Chanmueang (mix8645)