hamishivi/OLMo-1B-0724-SFT-hf

TEXT GENERATIONPricing:Input $0.04 / Cached $0.002 / Output $0.08Concurrent Unit Cost:1Model Size:1BQuant:BF16Context Size:32kPublished:Jul 17, 2024License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The hamishivi/OLMo-1B-0724-SFT-hf is a 1 billion parameter OLMo-based causal language model, fine-tuned for 5 epochs on the Tulu 2 dataset with a 32768 token context length. This model demonstrates improved performance across various benchmarks, including MMLU, GSM8k, and HumanEval, compared to its base version. It is particularly suited for general instruction-following tasks and serves as a strong foundation for further instruction tuning.

Loading preview...

Model Overview

The hamishivi/OLMo-1B-0724-SFT-hf is a 1 billion parameter OLMo-based language model, specifically the OLMo-1B-0724-hf base model, that has undergone Supervised Fine-Tuning (SFT). It was fine-tuned for 5 epochs using a learning rate of 1e-5 on a specific version of the Tulu 2 dataset, with a batch size of 1 and 128 gradient accumulation steps. The training utilized a linear warmup followed by linear decay.

Key Capabilities & Performance

This SFT model significantly improves upon its base model across several key benchmarks. It achieves notable gains in:

  • MMLU 0-shot: 36.0 (up from 25.0)
  • GSM8k CoT 8-shot: 12.5 (up from 7.0)
  • BBH CoT 3-shot: 27.2 (up from 22.5)
  • HumanEval P@10: 21.2 (up from 16.0)
  • TruthfulQA %Info+True: 40.9 (up from 23.0)
  • XSTest F1: 81.9 (up from 67.6)

It also shows a substantial reduction in toxicity, with a Toxigen % Toxic score of 59.7 compared to the base model's 80.3. The average performance across these metrics for this SFT model is 33.0, a significant improvement over the base model's 25.2.

Differentiators & Use Cases

This model stands out due to its targeted SFT on a high-quality instruction dataset, making it more capable at following instructions than its base counterpart. While a further DPO-trained 'Instruct' version exists with even higher overall performance, this SFT model offers a strong balance of performance and reduced toxicity. It is well-suited for general instruction-following tasks where a compact yet capable model is desired, and can serve as an excellent foundation for further specialized fine-tuning or as a robust general-purpose instruction-tuned model.

Finetunes
2 models