Orion-zhen/Qwen2.5-7B-Instruct-Uncensored

Model Overview

Orion-zhen/Qwen2.5-7B-Instruct-Uncensored is a 7.6 billion parameter language model derived from Qwen/Qwen2.5-7B-Instruct, developed by Orion-zhen. Its primary distinction is its uncensored nature, achieved through a fine-tuning process combining Supervised Fine-Tuning (SFT) and Direct Preference Optimization (DPO).

Key Capabilities & Training

This model was trained using a combination of SFT and DPO to ensure uncensored responses while preserving the original model's general capabilities. The SFT phase utilized datasets like NobodyExistsOnTheInternet/ToxicQAFinal and anthracite-org/kalo-opus-instruct-22k-no-refusal. For DPO, datasets such as Orion-zhen/dpo-toxic-zh, unalignment/toxic-dpo-v0.2, and Crystalcareai/Intel-DPO-Pairs-Norefusals were employed. The model supports a substantial context length of 131072 tokens and is multilingual, covering languages like Chinese, English, French, Spanish, and more.

Performance & Use Cases

Evaluations on the Open LLM Leaderboard show an average score of 27.99. Specific task results include 72.04 on IFEval (0-Shot) and 35.83 on BBH (3-Shot). While designed to be uncensored, the developer notes that it may still exhibit limitations in generating detailed descriptions for extreme scenarios, potentially due to pretraining data deletions in the base Qwen model. This model is suitable for applications where reduced refusal and more open-ended content generation are desired, particularly in multilingual contexts.