SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2310.05914 · 2023

NEFTune: Noisy Embeddings Improve Instruction Finetuning

arXiv · PDF · Open in the Atlas

Code that ran

We lifted 8 functions out of this paper's own repositories and ran 7 of them in a sandbox. "Ran" means the function executed on a synthesized input and returned a value. It is not a reproduction of the paper's results.

RepositoryRoleRan
neelsjain/neftune canonical 0 of 1
akjindal53244/arithmo pwc_unofficial 7 of 7
FunctionStatusWhere it lives
extract_ground_truth_answer Ran akjindal53244/arithmo/eval/gsm8k/gsm8k_compute_metric_zero_shot_CoT.py
code served (permissive licence) · get_code("466b972ae1ee4368")
extract_predcited_answer Ran akjindal53244/arithmo/eval/gsm8k/gsm8k_compute_metric_zero_shot_CoT.py
code served (permissive licence) · get_code("d5cb757080457d06")
extract_python_program Ran akjindal53244/arithmo/eval/gsm8k/gsm8k_write_zero_shot_PoT_outputs.py
code served (permissive licence) · get_code("bccdf81fed5534bd")
is_equiv Ran akjindal53244/arithmo/eval/MATH/MATH_compute_metric_zero_shot_CoT.py
code served (permissive licence) · get_code("20063c675d638a77")
last_boxed_only_string Ran akjindal53244/arithmo/eval/MATH/MATH_compute_metric_zero_shot_CoT.py
code served (permissive licence) · get_code("0b14c648516c38a7")
remove_boxed Ran akjindal53244/arithmo/eval/MATH/MATH_compute_metric_zero_shot_CoT.py
code served (permissive licence) · get_code("f3bbe264b05aadd3")
similar Ran akjindal53244/arithmo/data_prep/prepare_model_traininig_data.py
code served (permissive licence) · get_code("5e4b02e5bc753357")
generate_responses_batched Not yet run neelsjain/neftune/experiment_code/eval_generate.py
code served (permissive licence) · get_code("903c643febb1d508")

Repositories linked to this paper

Some links come from the archived Papers with Code dataset (CC BY-SA 4.0): attribution and licence.

Abstract

We show that language model finetuning can be improved, sometimes dramatically, with a simple augmentation. NEFTune adds noise to the embedding vectors during training. Standard finetuning of LLaMA-2-7B using Alpaca achieves 29.79% on AlpacaEval, which rises to 64.69% using noisy embeddings. NEFTune also improves over strong baselines on modern instruction datasets. Models trained with Evol-Instruct see a 10% improvement, with ShareGPT an 8% improvement, and with OpenPlatypus an 8% improvement. Even powerful models further refined with RLHF such as LLaMA-2-Chat benefit from additional training with NEFTune.

For agents

The same record, over MCP at https://syntology.ai/mcp:

get_harvested_code_for_paper("2310.05914")
get_code_for_paper("2310.05914")
have("2310.05914")

Connect an agent — have() is free.