nbeerbower/CHUD-Qwen3.6-27B
The nbeerbower/CHUD-Qwen3.6-27B is a 27 billion parameter language model, fine-tuned using the ORPO method on synthetic data derived from Grok 4.5. Built upon the nbeerbower/BigBubba-Qwen3.6-27B base model, it features a 32768 token context length. This model is specifically trained on a combination of DPO datasets, including 'weasel-dpo', 'seX-ai-dpo', and 'grok-politically-incorrect-dpo', indicating a focus on specific conversational styles or content generation. It is intended for use cases requiring a model with a particular training emphasis on these datasets.
Loading preview...
Model Overview
The nbeerbower/CHUD-Qwen3.6-27B is a 27 billion parameter language model, fine-tuned using the ORPO (Optimized Reward Prompting) method. It is built upon the nbeerbower/BigBubba-Qwen3.6-27B base model and was trained with a maximum sequence length of 2048 tokens over 2 epochs.
Training Details
This model was trained using the Merlina platform, employing a LoRA configuration with a rank of 32 and an alpha of 64. Key training parameters include a learning rate of 8e-06 and an effective batch size of 8. The training targeted specific modules: up_proj, down_proj, gate_proj, k_proj, q_proj, v_proj, and o_proj.
Datasets
The CHUD-Qwen3.6-27B was trained on a concatenation of three distinct DPO (Direct Preference Optimization) datasets:
These datasets, described as containing synthetic data from Grok 4.5, suggest a specialization in generating content aligned with the characteristics of these specific data sources.
Intended Use
Given its ORPO fine-tuning on specialized DPO datasets, this model is suitable for applications requiring responses or content generation that reflect the stylistic and thematic properties of its training data. Developers can reproduce the exact training configuration using the provided Merlina code snippet.