somieee20/llama-linkedin

Llama-3.1-8B-LinkedIn-Finetune

This model is a QLoRA fine-tuned version of Meta-Llama 3.1 8B Instruct, trained on a curated dataset of 6,200 viral-style LinkedIn posts inspired by top creators like Lara Acosta, Matt Gray, and Mischa G. It is optimized for high-conversion, authentic, and emotionally resonant content generation — particularly for founders, creators, and professionals looking to grow influence and inbound leads on LinkedIn.

🧠 Model Details

Model Description

• —Base model: meta-llama/Meta-Llama-3.1-8B-Instruct
• —Fine-tuning method: QLoRA (8-bit quantization, rank = 64)
• —Context length: 4096 tokens
• —Training samples: 6,200 (95 % train / 5 % eval)
• —Hardware: NVIDIA H100 80 GB GPU
• —Precision: bfloat16 with 8-bit loading
• —Epochs: 5
• —Optimizer: AdamW (Torch Implementation)
• —Learning rate: 3e-4 with cosine schedule
• —LoRA target modules: qproj, kproj, vproj, oproj, gateproj, upproj, downproj, lmhead

📊 Training Configuration

💾 Dataset

The dataset consists of 6000+ real viral posts from top creators, expanded synthetically to ~6,200 samples using generative augmentation for tone, structure, and narrative diversity. Each post follows the LinkedIn-native storytelling format (hook → story → lesson → CTA) with labeled stylistic attributes such as authenticity, pacing, and emotional tone.

⚙️ Use Cases

✅ Direct Use

• —Generating high-quality LinkedIn posts or storytelling templates.
• —Writing thought leadership content for founders or coaches.
• —Producing outbound copy for cold outreach that feels “human.”

🔧 Downstream Use

• —Integrate into marketing automation tools or content CRMs.
• —Use as a base for persona-tuned agents (e.g., “Matt-style writer”).
• —Fine-tune further for niche B2B verticals (e.g., SaaS, AI, VC).

🚫 Out-of-Scope

• —Political or sensitive opinion generation.
• —Automated spam or fake persona creation.

🧩 How to Use

from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "somieee20/llama-3.1-8b-linkedin"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)

prompt = "Write a viral LinkedIn post about learning from startup failures."
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
outputs = model.generate(**inputs, max_new_tokens=300, temperature=0.8, do_sample=True)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))