Fine-tuning blocked: “Unsafe training file”

API Bugs
, , ,
1

Hi,

I’m running into the following problem when trying to fine-tune a 4.1-mini (same for 4.1, 4.1-nano and all 4o models):

The job failed due to an unsafe training file. This training file was blocked by our moderation system because it contains too many examples that violate OpenAI’s usage policies, or because it attempts to create model outputs that violate OpenAI’s usage policies.

It’s a bit confusing because:

  • The exact same dataset fine-tunes on GPT-3.5 without any problem.

  • I checked the dataset on harmful content using the omni-moderation-latest model with a threshold of 0.001, and none of the entries were flagged.

  • The content is clean, fully policy-compliant, and contains no disallowed topics.

This seems to be a very common issue, since there are many other threads reporting the same “unsafe training file” error when fine-tuning GPT-4.1 models. Yet, so far, there hasn’t been any clear explanation from OpenAI on what specifically triggers these moderation blocks or how to fix them.

Could someone from the OpenAI team please take a look at my jobs or provide an explanation of the possible causes?

Here are the fine-tuning job IDs for reference:

  • ftjob-Yd4VFhq8AWGyRhhT1cWZOG9J (4.1-mini)

  • ftjob-As1ER1gva8rxaSU5KTTxGVkt (4o-mini)

Thanks in advance!

2 Likes
2

The safety blocks are awful

You can’t invest time & $ into crafting the optimal data set and params, only to have the same data set blocked a month later.

1 Like