CompassioninMachineLearning/Qwen3-8b-compassion-cleaned-10k-20260910-CPT-merged-epoch-4
Qwen3-8b-compassion-cleaned-10k-20260910-CPT-merged-epoch-4 is an 8 billion parameter Qwen3-based language model developed by CompassioninMachineLearning, featuring a 32,768 token context length. This model is a merged BF16 checkpoint from epoch 4.0, specifically trained on a cleaned dataset focused on compassion. It is designed for applications requiring nuanced understanding and generation related to compassionate themes, without requiring an adapter for loading.
Loading preview...
Model Overview
Qwen3-8b-compassion-cleaned-10k-20260910-CPT-merged-epoch-4 is an 8 billion parameter Qwen3-based model developed by CompassioninMachineLearning. This specific version is a standalone merged BF16 checkpoint from epoch 4.0, step 1500, and does not require an adapter for loading.
Training Details
The model was trained using the CompassioninMachineLearning/compassion_12185_cleaned dataset. The training regimen involved 10,000 distinct documents per epoch, with an additional 2,000 repeat exposures, and 200 disjoint validation documents. The weights are validated BF16 and are packaged losslessly into eight safetensors shards.
Key Characteristics
- Base Model: Qwen3-8b
- Parameter Count: 8 billion
- Context Length: 32,768 tokens
- Training Focus: Compassion-related content, utilizing a specifically curated and cleaned dataset.
- Merged Checkpoint: This is a fully merged model, ready for direct use without additional adapters.
Important Note
The developers explicitly state that the training process itself "does not establish an improvement in compassion"; evaluation of its compassionate capabilities should be conducted separately. Further details on the base revision, document selection hashes, training parameters, and export validation can be found in the run_manifest.json file.