Model Overview

Orion-zhen/Llama3-70B-Orion-Chinese is a 70 billion parameter language model derived from failspy/Llama-3-70B-Instruct-abliterated-v3 through DPO (Direct Preference Optimization) fine-tuning. A key differentiator of this model is its deliberate retention of Llama 3's lively tone and propensity for emoji usage, setting it apart from other Chinese Llama 3 adaptations. The model has a context length of 8192 tokens, with potential for extension up to 32k tokens by modifying max_position_embeddings and rope_theta in the config.json.

Key Capabilities

• Chinese Language Proficiency: Optimized for conversational and reasoning tasks in Chinese.
• Lively Tone & Emoji Usage: Retains the characteristic Llama 3 style, making interactions more engaging.
• Logical Reasoning: Demonstrates strong performance in complex logical and mathematical problem-solving.
• Code Generation: Capable of generating code, as shown in examples for LeetCode problems.
• Safety & Ethics: Includes examples of refusal to engage with unsafe prompts, indicating some safety alignment.

Training Details

The model was trained using LLaMA-Factory with bitsandbytes 4-bit QLoRA. The training leveraged a diverse set of DPO datasets, including unalignment/toxic-dpo-v0.2, shareAI/DPO-zh-en-emoji, hiyouga/DPO-En-Zh-20k, argilla/distilabel-math-preference-dpo, antiven0m/physical-reasoning-dpo, AlekseyKorshuk/evol-codealpaca-v1-dpo, and Intel/orca_dpo_pairs. Key hyperparameters included a learning rate of 1e-05, a total batch size of 8, and 3 epochs of training.