jekunz's picture
Update README.md
3a80f16 verified
metadata
license: apache-2.0
datasets:
  - HuggingFaceFW/fineweb-2
language:
  - sv
base_model:
  - HuggingFaceTB/SmolLM2-135M-Instruct
pipeline_tag: text-generation

Model created for the paper "Preferences for Idiomatic Language are Acquired Slowly --- and Forgotten Quickly: A Case Study on Swedish", TACL 2026.

Citation

@misc{kunz2026preferencesidiomaticlanguageacquired,
      title={Preferences for Idiomatic Language are Acquired Slowly -- and Forgotten Quickly: A Case Study on Swedish}, 
      author={Jenny Kunz},
      year={2026},
      eprint={2602.03484},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2602.03484}, 
}

Training:

This is a SmolLM2-135M model continually pre-trained on the Swedish portion of Fineweb-2.

  • 1 Epoch
  • Learning rate: 5e-4
  • LR scheduler: Cosine
  • Warmup ratio: 0.05
  • Batch size: 1
  • 4 A100 (40GB) GPUs
  • Gradient accumulation steps: 64
  • Effective batch size: 256
  • Max. context length: 8192 tokens

Limitations

This is a research model intended for studying pre-training dynamics and I do not recommend using it for any practical purposes. It is trained on a web corpus, and no alignment whatsoever has been performed, which means that the model will likely reflect its training data's biases and produce lots of hallucinations.