LLMWriter-8B

LLMWriter-8B is a fine-tuned version of Llama 3.1 8B designed to improve writing quality, instruction following, content generation, and structured response creation.

• Developed by: aldenirsrv
• License: Apache 2.0
• Finetuned from model: unsloth/Llama-3.1-8B

This model was trained using LoRA (Low-Rank Adaptation) with the Unsloth framework and Hugging Face TRL.

This Llama model was trained 2x faster with Unsloth and Hugging Face's TRL library.

---

Model Description

LLMWriter-8B is focused on:

• Content creation
• Technical writing
• Documentation generation
• Blog writing
• Social media content
• Instruction following
• Writing assistance
• General-purpose text generation

The model was trained on a curated instruction-following dataset containing over 10,000 examples.

---

Training Dataset

DatasetDict({
    train: Dataset({
        features: ['instruction', 'output', 'source', 'score'],
        num_rows: 10384
    })
    test: Dataset({
        features: ['instruction', 'output', 'source', 'score'],
        num_rows: 547
    })
})

Split	Samples
Train	10,384
Test	547
Total	10,931

---

Training Configuration

Parameter	Value
Base Model	Llama 3.1 8B
Fine-Tuning Method	LoRA
Framework	Unsloth
Epochs	3
GPUs	1
Batch Size	2
Gradient Accumulation	8
Effective Batch Size	16
Trainable Parameters	83,886,080
Total Parameters	8,114,147,328
Percentage Trained	1.03%

Training Summary:

Num examples = 10,384
Num Epochs = 3
Total steps = 1,947

Trainable parameters:
83,886,080 of 8,114,147,328
(1.03% trained)

---

Training Results

Optimization Metrics

Metric	Value
Initial Loss	1.2151
Final Loss	0.2187
Average Training Loss	0.5334
Initial Learning Rate	3e-4
Final Learning Rate	1.54e-7
Min Gradient Norm	0.1587
Max Gradient Norm	2.3887

Key Observations

✅ Consistent loss reduction throughout training

✅ Stable gradient norms with no gradient explosion

✅ Effective learning-rate decay schedule

✅ Smooth convergence after three epochs

✅ Stable LoRA fine-tuning process

---

Performance

Metric	Value
Training Time	~1h41m
Total FLOPs	~7.66e17
Samples per Second	~5.1
Steps per Second	~0.32
Time per Step	~3.1s

The training run completed successfully and demonstrated stable convergence without optimization instability.

---

Intended Use

LLMWriter-8B is intended for:

• Writing assistance
• Content generation
• Blog creation
• Documentation drafting
• Technical writing
• Knowledge articles
• Social media posts
• Structured responses
• General instruction-following tasks

---

Example Prompt

Write a professional LinkedIn post explaining why Small Language Models (SLMs) are becoming important for enterprise AI adoption.

---

Usage

vLLM

vllm serve aldenirsrv/LLMWriter-8B \
  --host 0.0.0.0 \
  --port 8888

Transformers

from transformers import AutoTokenizer, AutoModelForCausalLM

model_name = "aldenirsrv/LLMWriter-8B"

tokenizer = AutoTokenizer.from_pretrained(model_name)

model = AutoModelForCausalLM.from_pretrained(
    model_name,
    device_map="auto"
)

prompt = "Write a blog post introduction about AI governance."

inputs = tokenizer(
    prompt,
    return_tensors="pt"
).to(model.device)

outputs = model.generate(
    **inputs,
    max_new_tokens=300,
    temperature=0.7
)

print(
    tokenizer.decode(
        outputs[0],
        skip_special_tokens=True
    )
)

OpenAI-Compatible API (vLLM)

from openai import OpenAI

client = OpenAI(
    base_url="http://localhost:8888/v1",
    api_key="dummy"
)

response = client.chat.completions.create(
    model="aldenirsrv/LLMWriter-8B",
    messages=[
        {
            "role": "user",
            "content": "Write a blog post about AI governance."
        }
    ]
)

print(response.choices[0].message.content)

---

Limitations

• The model may generate inaccurate information.
• Outputs should be reviewed before publication.
• Not intended for legal, medical, or financial advice.
• Performance depends on prompt quality and task complexity.
• The model has not been evaluated on standardized benchmark suites.

---

Future Work

Planned improvements include:

• Human preference evaluation
• Benchmark comparisons against the base model
• Additional instruction tuning
• Domain-specific fine-tuning
• Expanded evaluation datasets
• Quantized deployment variants

---

Acknowledgements

This model was fine-tuned using:

• Unsloth
• Hugging Face Transformers
• Hugging Face TRL
• PEFT (LoRA)
• PyTorch
• Comet ML

Special thanks to the teams behind Llama, Hugging Face, and Unsloth for enabling efficient open-source model development.

---

Author

Aldenir Flauzino

Software Engineer specializing in:

• Distributed Systems
• Platform Engineering
• AI Infrastructure
• Retrieval-Augmented Generation (RAG)
• Multi-Agent Systems
• Production AI Platforms

GitHub: https://github.com/aldenirsrv

Website: https://aldenir.me