Hummingbird

🎉 You are looking at Hummingbird 2, trained on a much more efficient corpus, achieving similar performance with 3x less parameters!

Hummingbird is a GPT-2 derivative trained to be conversational.

Training

The model was trained using the paged_adamw_8bit optimizer, gradient checkpointing, 500 steps, 1 batch size, and 4 gradient accumulation steps.

Datasets

The training corpus is made up of:

• First 1400 rows of qikp/reborn-5k-no-thoughts
• First 500 rows of HuggingFaceTB/smol-smoltalk
• First 100 rows of HuggingFaceTB/everyday-conversations-llama3.1-2k

The train / train_sft splits were used.

Chat template

The Zephyr chat template was used.

Limitations

The model frequently outputs incorrect information, confirmation with a larger, mature model is advised.

Benchmark

This model was tested against GAIA and compared using embeddings. See the results here.