Remove GPT-4 as the default model. #1072
Conversation
Codecov ReportAttention:
Additional details and impacted files@@ Coverage Diff @@
## main #1072 +/- ##
===========================================
+ Coverage 30.71% 62.88% +32.17%
===========================================
Files 30 30
Lines 4047 4047
Branches 916 965 +49
===========================================
+ Hits 1243 2545 +1302
+ Misses 2724 1264 -1460
- Partials 80 238 +158
Flags with carried forward coverage won't be shown. Click here to find out more. ☔ View full report in Codecov by Sentry. |
rickyloynd-microsoft
left a comment
rickyloynd-microsoft
left a comment
There was a problem hiding this comment.
Makes sense.
|
@rickyloynd-microsoft It's interesting... the compressible agent tests are failing now because they are reading a "model" attribute from llm_config rather than from the config_list. That's a bug with compressible agents, but previously hidden with this default! |
Turn on the lights and the bugs are revealed! |
…wise default to gpt-4. Revist after #1073 is addressed.
…nfig_list was always empty!
Co-authored-by: Chi Wang <wang.chi@microsoft.com>
sonichi
left a comment
sonichi
left a comment
There was a problem hiding this comment.
Approving despite the test failure: https://github.com/microsoft/autogen/actions/runs/7418135849/job/20185641305?pr=1072
because it's due to a design flaw in CompressibleAgent: #1073
* Remove GPT-4 as the default model. * Updated test_compressible_agent to work around a bug that would otherwise default to gpt-4. Revist after microsoft#1073 is addressed. * Worked around another bug in test_compressible_agent. It seems the config_list was always empty! * Reverted changes to compressible agent. * Noted that GPT-4 is the preferred model in the OAI_CONFIG_LIST_sample and README. * Fixed failing tests after microsoft#1110 * Update OAI_CONFIG_LIST_sample Co-authored-by: Chi Wang <wang.chi@microsoft.com> --------- Co-authored-by: Chi Wang <wang.chi@microsoft.com>
Why are these changes needed?
Defaulting to GPT-4 can lead to unexpected costs -- especially when the default kicks due to typos or simple misconfigurations. This removes the default, causing model misconfigurations to raise errors.
Related issue number
Addresses #1025
Checks