ministral-8B-texthumanizer

TextHumanizer - Fine-tuned Ministral-8B-Instruct:

TextHumanizer is a fine-tuned version of the Ministral-8B-Instruct model. It is designed to transform robotic, overly-formal, or synthetic AI-generated text into fluent, natural, and human-like language.

This model was fine-tuned using Apple's MLX framework on a custom dataset of AI-generated text paired with humanized rewrites.

---

Model Details:

• Base model: mistralai/Ministral-8B-Instruct
• Model size: 8B parameters
• Fine-tuned using: MLX (Apple Silicon-native training)
• Fine-tuning method: QLora
• Precision: float16
• Hardware used: Apple Silicon M1 Macbook Pro
• Training time: 10 mins
• Epochs / Steps: 200
• Batch size: 4

---

Training Metrics:

---

Dataset:

• Source: Custom synthetic-to-human text pairs
• Size: 10k rows
• Structure: Pairs of (synthetic_input, humanized_output)
• Preprocessing: Standard MLX tokenization, formatting for instruct tuning
• License: Public

---

Capabilities:

TextHumanizer is designed to:

• Improve fluency and tone of AI outputs
• Make answers sound more relatable and natural
• Polish robotic or overly formal language

---

MLX (Apple Silicon):

from mlx_lm import load, generate

model, tokenizer = load("vishal-adithya/ministral-8B-texthumanizer")
prompt = "Instruction: Make this sound more natural.Input: The individual proceeded to consume nourishment.Response:"
generate(model, tokenizer, prompt)