From b2e3afec5b6c5e6b309bbfd30378a215bcb4f758 Mon Sep 17 00:00:00 2001 From: oliveradk Date: Fri, 26 Jul 2024 13:54:18 -0700 Subject: [PATCH 1/2] fix ioi paths and check add_bos_token on tokenizer --- auto_circuit/tasks.py | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-) diff --git a/auto_circuit/tasks.py b/auto_circuit/tasks.py index 40e6503..810321e 100644 --- a/auto_circuit/tasks.py +++ b/auto_circuit/tasks.py @@ -205,6 +205,8 @@ def init_task(self): bs_2 = bs[1] if isinstance(bs, tuple) else bs count_2 = b_count[1] if isinstance(b_count, tuple) else b_count has_tokenizer = hasattr(model, "tokenizer") and model.tokenizer is not None + if has_tokenizer and hasattr(model.tokenizer, "add_bos_token"): + model.tokenizer.add_bos_token = False train_loader, test_loader = load_datasets_from_json( model=model if has_tokenizer else None, path=repo_path_to_abs_path(f"datasets/{self._dataset_name}.json"), @@ -296,7 +298,7 @@ def init_task(self): key="Indirect Object Identification Component Circuit", name="Indirect Object Identification", _model_def="gpt2-small", - _dataset_name="ioi_prompts", + _dataset_name="ioi/ioi_prompts", batch_size=64, batch_count=2, _true_edge_func=ioi_true_edges, @@ -306,7 +308,7 @@ def init_task(self): key="Indirect Object Identification GPT2 Autoencoder Component Circuit", name="Indirect Object Identification", _model_def="gpt2-small", - _dataset_name="ioi_prompts", + _dataset_name="ioi/ioi_prompts", batch_size=1, batch_count=2, _true_edge_func=None, From 03b045bcb65f11ae937c1142a8c51afd75dc7ec8 Mon Sep 17 00:00:00 2001 From: Oliver Daniels-Koch <40397426+oliveradk@users.noreply.github.com> Date: Thu, 22 Aug 2024 17:39:31 -0400 Subject: [PATCH 2/2] comment on add_bos_token --- auto_circuit/tasks.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/auto_circuit/tasks.py b/auto_circuit/tasks.py index 810321e..876f47c 100644 --- a/auto_circuit/tasks.py +++ b/auto_circuit/tasks.py @@ -205,7 +205,7 @@ def init_task(self): bs_2 = bs[1] if isinstance(bs, tuple) else bs count_2 = b_count[1] if isinstance(b_count, tuple) else b_count has_tokenizer = hasattr(model, "tokenizer") and model.tokenizer is not None - if has_tokenizer and hasattr(model.tokenizer, "add_bos_token"): + if has_tokenizer and hasattr(model.tokenizer, "add_bos_token"): # ensures bos token not added to answers/wrong answers for pythia models model.tokenizer.add_bos_token = False train_loader, test_loader = load_datasets_from_json( model=model if has_tokenizer else None,