GhostWriterLlama-3.2-1B

• Developed by: Ahmed Shahriar Sakib
• License: Apache 2.0
• Finetuned from model: unsloth/Llama-3.2-1B
• Fine-tuning dataset: ahmedshahriar/llmGhostWriter (instruction-response)
• Use-case: Writing/ghost-writing style assistant for generating expository and creative text

Description

GhostWriterLlama-3.2-1B is an instruction-tuned variant of Llama 3.2 (1B) optimized for writing-style tasks, developed using the Unsloth framework for faster training and efficient memory usage. The model is tailored for text generation in English and supports use cases like blog drafts, creative writing, and content assistance.

Training Details

• Fine-tuning framework: Unsloth + Hugging Face TRL (as referenced in the base model documentation)
• Data: Derived from the llmGhostWriter dataset of ~2.1K instruction-response pairs, curated from publicly available articles and refined via LLM assistance and manual review.
• Objective: Supervised fine-tuning (SFT) on writing/ghost-writing tasks.
• Infrastructure & Resources: This llama model was trained 2x faster with Unsloth and Huggingface's TRL library.
• Context length & quantization: Inherits the capabilities of the base model (1B params) and supports standard inference via the Transformers ecosystem.

Intended Use

This model is designed for text-generation applications where users require coherent, stylistically consistent writing outputs. Typical use cases include drafting articles, blog posts, marketing copy, and assisting in creative writing.

Out-of-Scope and Risks

• Do not rely on this model for factual retrieval, high-stakes decision making, or tasks requiring verified accuracy—outputs may contain errors or misleading claims.
• The instruction tuning was done on a limited (~2K example) dataset of curated writing tasks, so the model may inherit stylistic biases or lack broad domain coverage.
• It is not explicitly aligned for multi-turn chat or tool-use—inputs outside the training distribution may degrade performance.

Evaluation

Evaluated with an LLM-as-a-judge setup on the test split of the ahmedshahriar/llmGhostWriter dataset. A judge model (GPT-4.1-nano) scored each generated response on a 1–3 scale for two criteria:

• Accuracy — factual correctness and completeness
• Style — appropriateness of tone for blog/social content (non-academic)

Quantitative averages from these scores, combined with qualitative review, indicated similar factual accuracy and slightly improved stylistic alignment over the base Llama-3.2-1B-Instruct in ghostwriting/expository prompts. (Caveat: LLM-judge metrics can reflect the judge’s biases; consider complementary human review for critical use.)

For full training and evaluation workflows, see the Colab notebooks:

• Fine-tuning Notebook
  

• Evaluation Notebook
  

Evaluation results (model outputs and judge scores) are publicly available for transparency at: ahmedshahriar/GhostWriterLlama-3.2-1B-results. The dataset includes model generations and GPT-4.1-nano evaluation scores on the test split of the ahmedshahriar/llmGhostWriter dataset.

Limitations

• Scope of training data (~2,100 examples) is modest; topic and style diversity may be limited.
• Not trained via RLHF or preference optimization (this is addressed in the subsequent DPO-tuned model).
• English only; performance on other languages is untested.

Citation

@misc{ahmedshahriar_ghostwriterllama3_2_1b_2025,
  author       = {Ahmed Shahriar Sakib},
  title        = {GhostWriterLlama-3.2-1B},
  year         = {2025},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/ahmedshahriar/GhostWriterLlama-3.2-1B}},
  license      = {Apache-2.0}
}

Acknowledgements

Thanks to the Unsloth team for the efficient fine-tuning framework, and to the authors of the publicly available articles used for dataset creation.