CompassioninMachineLearning/Qwen3-8b-compassion-seed2-20261001-CPT-merged-epoch-1
Qwen3-8b-compassion-seed2-20261001-CPT-merged-epoch-1 is an 8 billion parameter language model developed by CompassioninMachineLearning, based on the Qwen3 architecture. This model was trained on the 'compassion_12185_cleaned' dataset with 10,000 distinct documents and 2,000 repeat exposures per epoch, and has a context length of 32768 tokens. It is provided as a standalone BF16 merged model, requiring no adapter for loading, and is specifically designed for evaluating compassion-related tasks.
Loading preview...
Overview
Qwen3-8b-compassion-seed2-20261001-CPT-merged-epoch-1 is an 8 billion parameter language model from CompassioninMachineLearning, built upon the Qwen3 architecture. This model is a standalone BF16 merged version, meaning it can be loaded directly without the need for an adapter. It was trained for one epoch, specifically at step 375, and is packaged losslessly into eight safetensors shards.
Key Capabilities
- Standalone Operation: Provided as a merged BF16 model, eliminating the need for external adapters.
- Specialized Training: Trained on the
CompassioninMachineLearning/compassion_12185_cleaneddataset, focusing on compassion-related content. - Robust Merging: Utilizes Unsloth's native
save_pretrained_mergedmethod for a validated BF16 output.
Good For
- Compassion Evaluation: Designed for tasks involving the evaluation of compassion, though the README notes that training does not guarantee an improvement in compassion and requires separate evaluation.
- Research on Specialized Datasets: Ideal for researchers working with specific, curated datasets like
compassion_12185_cleaned. - Direct Deployment: Suitable for users who prefer a ready-to-use, merged model without additional setup for adapters.