Model Card for Qwen2.5-3B-Unlimited-Beta-200

Model Details

Model Description

This model is a fine-tuned version of the Qwen2.5-3B-Instruct model, optimized to remove response restrictions using a custom dataset. It employs Parameter-Efficient Fine-Tuning (PEFT) with LoRA (Low-Rank Adaptation) to enhance flexibility and performance on unrestricted topics, including sensitive and controversial queries. The model excels in natural language processing tasks, delivering creative and precise responses.

• Developed by: Ukiuky Tech
• Funded by: Ukiuky Tech
• Shared by: Ukiuky Tech
• Model type: Causal Language Model
• Language(s) (NLP): English and Chinese
• License: Apache 2.0
• Finetuned from model: Qwen/Qwen2.5-3B-Instruct

Model Sources

Uses

Direct Use

The model is designed for generating unrestricted responses to a wide range of queries, including sensitive and controversial topics. It is suitable for applications like chatbots, research tools, or creative writing assistants where flexibility and accuracy are critical.

Downstream Use

This model can be further fine-tuned for specialized applications, such as open-domain question-answering systems, content generation platforms, or educational tools requiring unrestricted responses.

Notes: We do not impose restrictions on the use of this model. Please comply with local laws and regulations.

Bias, Risks, and Limitations

The model inherits biases from the Qwen2.5-3B-Instruct base model and the custom dataset. Potential risks include:

• Amplifying biases in sensitive topics if the dataset is not comprehensive.
• Limited generalization to languages or domains outside the training scope.
• Potential for inappropriate content generation without proper prompting.

Recommendations

Users should validate outputs, monitor for biases, and clearly communicate the model’s limitations to prevent misuse.

Training Details

Training Data

The model was fine-tuned on a custom dataset of instructions and responses focused on unrestricted topics, enabling open-ended and creative dialogue.

Training Procedure

Preprocessing

The dataset was tokenized with a maximum length of 128 tokens, using a structured prompt format: <|im_start|>system\n{instruction}<|im_end|>\n<|im_start|>user\n{input}<|im_end|>\n<|im_start|>assistant\n{output}<|im_end|>. Labels ignored prompt tokens for loss calculation.

Training Hyperparameters

• Training regime: bf16 mixed precision
• Batch size: 8 (per device, gradient accumulation steps of 2)
• Learning rate: 5e-5
• Epochs: 1
• Optimizer: AdamW
• Gradient clipping: Max norm of 1.0
• LoRA configuration: Rank (r): 8, Alpha: 32, Dropout: 0.1, Target modules: ["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"]
• Gradient checkpointing: Enabled

Speeds, Sizes, Times

Training required approximately 1000 GPU hours. Recommended hardware: 4 NVIDIA RTX5090s.

Evaluation

The model was evaluated on a test split of the custom dataset, focusing on response accuracy and appropriateness for unrestricted topics.

Technical Specifications

Model Architecture and Objective

A transformer-based causal language model with 3 billion parameters, fine-tuned with LoRA to maximize response flexibility.

Compute Infrastructure

Hardware: Not publicly disclosed. Software: Transformers (Latest), PEFT 0.15.0, PyTorch (Latest), SwanLab, Datasets.

Model Card Authors

Ukiuky Tech

Model Card Contact

Email: [email protected]