At some point, I dug into this as well, realizing that lower scores do not necessarily mean negative, and higher scores do not necessarily mean positive. Could you prepare for a snippet and run for a few more examples?

Here's a snippet for running the output data in the notebook through the pipeline:

df = pd.read_csv(io.StringIO('''
query,response (after),rewards_src
"Oh dear,",I must say that I are hanging my head on this,-1.007609
I've seen,"three million dialogue throughout, and",2.240883
Hi:<br /><br,/>I also like that movie. It's so funny,2.415630
I'm a writer,", not a screenwriter. I've written",-0.724324
If you,"are looking at the cinematography, the acting,",0.148751
OMG this,"movie was totally wonderful, I it was the ide...",2.590190
'''), sep=",")
texts = [q + r for q, r in zip(df["query"], df["response (after)"])]

sent_kwargs = {"top_k": None, "function_to_apply": "none", "batch_size": 16}
sentiment_pipe = pipeline("sentiment-analysis", model="lvwerra/distilbert-imdb")

old_method = [output[1]["score"] for output in sentiment_pipe(texts, **sent_kwargs)]

pipe_outputs = sentiment_pipe(texts, **sent_kwargs)
new_method = [item["score"] for output in pipe_outputs for item in output if item["label"] == "POSITIVE"]

pd.DataFrame({'source_reward': df['rewards_src'], 'old_reward': old_method, 'new_reward': new_method})

source_reward (the one pre-computed in the notebook) matches the method I committed but not the one currently in the notebook. And from my experimentation the sent pipeline seems to give expected results.

You can see that the existing method just selects the wrong class sometimes, by removing the ["score"]

I still have the results from full training runs I did on both versions of the notebook:

Using the existing notebook (both training and eval metric were wrong):

And the new one:

mean:
rewards (before)    0.239116
rewards (after)     2.475334
dtype: float64
median:
rewards (before)    0.422371
rewards (after)     2.701585
dtype: float64

I don't have the checkpoints anymore, but the jump in positive sentiment is in line with what would be expected (and roughly matches the results saved in the notebook)

@vwxyzjn I refreshed the output table using the new results I got; I think they show the delta between the models pretty well. Also, here are the logs for the training run: https://wandb.ai/cemiu-team/trl-gpt2-sentiment/runs/01ch6t6i

FYI, the wandb.ai link in the notebook is not publicly accessible. Someone with admin access should change the permissions.

This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.

@vwxyzjn have you had a change to look at this since?

Thanks for the PR! The wandb link has been updated.

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.