SN-13B-8k-Instruct

Basic Information

• Blog Post: Link
• Discord: Link
• SN-13B-8k-Instruct: Link

Reproducibility Information

This repo contains the reproducibility information for the numbers listed in the SN-13B-8k-Instruct blogpost. Scrolls and ZeroScrolls refer to the following benchmarks:

1. Scrolls Benchmark
2. ZeroScrolls Benchmark

Setup Eleuther AI LM Evaluation Harness

1. git clone https://github.com/EleutherAI/lm-evaluation-harness.git
2. Checkout the commit of LM Evaluation Harness that we used to collect the results:

git checkout fe803c2920a85f6afb74ea05d1d2f98ec27f1a63`

3. Follow the setup instructions specified in the repository's README.

ZeroScrolls Reproducibility

1. Add ZeroScrolls task code to the LM Evaluation Harness.
   • This will involve importing the zero scrolls tasks in the tasks/__init__.py file in LM Evaluation Harness. You will need to add the following line to the TASK_REGISTRY:

   **zero_scrolls.construct_tasks(),

2. Install requirements

pip install requirements.txt

3. Run the following command in the LM Evaluation Harness:

python main.py --batch_size 1 --tasks zero_scrolls_gov_report,zero_scrolls_summ_screen_fd,zero_scrolls_qm_sum,zero_scrolls_squality,zero_scrolls_qasper,zero_scrolls_narrative_qa,zero_scrolls_quality,zero_scrolls_musique,zero_scrolls_space_digest,zero_scrolls_book_sum_sort --model gpt2 --model_args pretrained=sambanovasystems/SN-13B-8k-Instruct,dtype=float16 --num_fewshot 0 --no_cache

Scrolls Reproducibility

1. In the LM Evaluation Harness, open tasks/scrolls.py and replace the '\n' with your model's end of text token in the until list for all greedy_until requests.
2. Run the following command in the LM Evaluation Harness:

python main.py --batch_size 1 --tasks scrolls_govreport,scrolls_qmsum,scrolls_quality,scrolls_summscreenfd --model gpt2 --model_args pretrained=sambanovasystems/SN-13B-8k-Instruct,dtype=float16 --num_fewshot 0  --no_cache