Test network using differential attention instead of classical attention (using nope). Other than some alterations to the attention, this is otherwise the same configuration as https://huggingface.co/HuggingFaceTB/SmolLM2-135M-Instruct

Scripts:

inference.py to run the model with some test prompts
test_train.py runs with the exact configurations used to train this model and is the reproduction script. Data is assumed to be in JSONL format with "text":"example text", "text":"..."

Notes:

Compared to the control model of Smollm2, this is bordering on incoherent. Potentially this model size is too small to correctly leverage differential attention. It's clearly picked up on some ideas in language, but is generally worse than the control model using GQA in terms of human output.

Training Metrics

Dataset Information

Training data per epoch: 1 GB
Total tokens trained: 48,261,120
No sythetic data

Training Results

Final Train Loss: 2.6883
Final Train Perplexity: 14.71