From 10c2f6e360d60730ab6595ecab2947c0c5aaf2ee Mon Sep 17 00:00:00 2001 From: Nina Xu Date: Fri, 23 May 2025 19:02:35 +0000 Subject: [PATCH] change default rope to auto for v2 configs --- .../gretel/tasks/tabular_ft__default.yaml | 11 ++++++----- .../tasks/tabular_ft__differential_privacy.yaml | 11 ++++++----- config_templates/gretel/workflows/tabular-ft-dp.yml | 11 ++++++----- config_templates/gretel/workflows/tabular-ft.yml | 11 ++++++----- 4 files changed, 24 insertions(+), 20 deletions(-) diff --git a/config_templates/gretel/tasks/tabular_ft__default.yaml b/config_templates/gretel/tasks/tabular_ft__default.yaml index 2b25d20e..eae866a3 100644 --- a/config_templates/gretel/tasks/tabular_ft__default.yaml +++ b/config_templates/gretel/tasks/tabular_ft__default.yaml @@ -30,11 +30,12 @@ task: # Scale the base LLM's context length by this factor # using RoPE scaling to handle datasets with more # columns, or datasets containing groups with more - # than a few records. You can try increasing the - # rope_scaling_factor (you could first try the value 2) - # if you hit an error for maximum tokens. It must be - # an integer value. The default is 1 and maximum is 6. - rope_scaling_factor: 1 + # than a few records. If set to 'auto', we will + # estimate a value that's enough to cover your + # dataset. Try increasing this value if you hit an + # error for maximum tokens. It must be an integer + # value between 1 and 6. + rope_scaling_factor: auto generate: num_records: 1000 diff --git a/config_templates/gretel/tasks/tabular_ft__differential_privacy.yaml b/config_templates/gretel/tasks/tabular_ft__differential_privacy.yaml index fbc15fd2..01044be6 100644 --- a/config_templates/gretel/tasks/tabular_ft__differential_privacy.yaml +++ b/config_templates/gretel/tasks/tabular_ft__differential_privacy.yaml @@ -38,11 +38,12 @@ task: # Scale the base LLM's context length by this factor # using RoPE scaling to handle datasets with more # columns, or datasets containing groups with more - # than a few records. You can try increasing the - # rope_scaling_factor (you could first try the value 2) - # if you hit an error for maximum tokens. It must be - # an integer value. The default is 1 and maximum is 6. - rope_scaling_factor: 1 + # than a few records. If set to 'auto', we will + # estimate a value that's enough to cover your + # dataset. Try increasing this value if you hit an + # error for maximum tokens. It must be an integer + # value between 1 and 6. + rope_scaling_factor: auto # You can try increasing this until you run out-of-memory. batch_size: 4 diff --git a/config_templates/gretel/workflows/tabular-ft-dp.yml b/config_templates/gretel/workflows/tabular-ft-dp.yml index 9672ea67..fc197491 100644 --- a/config_templates/gretel/workflows/tabular-ft-dp.yml +++ b/config_templates/gretel/workflows/tabular-ft-dp.yml @@ -46,11 +46,12 @@ steps: # Scale the base LLM's context length by this factor # using RoPE scaling to handle datasets with more # columns, or datasets containing groups with more - # than a few records. You can try increasing the - # rope_scaling_factor (you could first try the value 2) - # if you hit an error for maximum tokens. It must be - # an integer value. The default is 1 and maximum is 6. - rope_scaling_factor: 1 + # than a few records. If set to 'auto', we will + # estimate a value that's enough to cover your + # dataset. Try increasing this value if you hit an + # error for maximum tokens. It must be an integer + # value between 1 and 6. + rope_scaling_factor: auto # You can try increasing this until you run out-of-memory. batch_size: 4 diff --git a/config_templates/gretel/workflows/tabular-ft.yml b/config_templates/gretel/workflows/tabular-ft.yml index 05d4a008..3839281a 100644 --- a/config_templates/gretel/workflows/tabular-ft.yml +++ b/config_templates/gretel/workflows/tabular-ft.yml @@ -38,11 +38,12 @@ steps: # Scale the base LLM's context length by this factor # using RoPE scaling to handle datasets with more # columns, or datasets containing groups with more - # than a few records. You can try increasing the - # rope_scaling_factor (you could first try the value 2) - # if you hit an error for maximum tokens. It must be - # an integer value. The default is 1 and maximum is 6. - rope_scaling_factor: 1 + # than a few records. If set to 'auto', we will + # estimate a value that's enough to cover your + # dataset. Try increasing this value if you hit an + # error for maximum tokens. It must be an integer + # value between 1 and 6. + rope_scaling_factor: auto generate: num_records: 1000