only a change in case of a letter changes the output of openAI GPT2Model

[Jessen-Li]

In cha05.ipynb, the last code block

torch.manual_seed(123)

token_ids = generate(
    model=gpt,
    idx=text_to_token_ids("Every effort moves you", tokenizer).to(device),
    max_new_tokens=25,
    context_size=NEW_CONFIG["context_length"],
    top_k=50,
    temperature=1.5
)

print("Output text:\n", token_ids_to_text(token_ids, tokenizer))

And the following is the output of my MacBook:

Output text:
 Every effort moves you toward finding an ideal new way to practice something!

What makes us want to be on top of that?

Only when I change effort to Effort, that is

idx=text_to_token_ids("Every Effort moves you", tokenizer).to(device),

The output becomes, (the first line the same, the second line varies significantly)

Output text:
 Every Effort moves you toward finding an ideal new way to practice something!

What makes this list unique is that, because of their limited selection

Does this mean openAI GPT2 model is highly untrustworthy?

[casinca]

Hello,

The difference between "effort" and "Effort" may seem meaningless but behind the scenes, they are tokenized differently, and since you feed different tokens to the LLM, you'll get a different trajectory. You can try by yourself on https://tiktokenizer.vercel.app/?model=gpt2 and see the difference in tokenization between the 2 sentences in lowercase/capitalized.

It''s not just GPT2, as long as they are tokenized differently, you can try also on Google AI studio, with a non reasoning model like GEMMA, set temperature to 0, feed the lowercase sentence, rerun and you'll see no difference. Rerun with capitalized, you'll see a slightly different answer but close semantically.

Although i don't have a good answer as to why only/specifically the 2nd part differs in this case.

[d-kleine]

Unfortunately, @rasbt is currently recovering from an injury, but he will take a look into this question as soon as he's feeling better.

Although i don't have a good answer as to why only/specifically the 2nd part differs in this case.

Yeah, same. I suspect it might be related to the mps issue possibly? My output is:

Output text:
 Every Effort moves you as far as the target can go until the end of your turn. By doing so you'll ensure those you're defending gain

[casinca]

Sorry for the late reply. Yeah it's more likely to get a somewhat different trajectory than the same exact full answer, token for token. Even though it's still possible but very unlikely. Maybe the fact that GPT2 is small with less expressive power, or a short momentum effect for choosing the first same word, which led to choosing the same 2nd word etc... or simply luck. These are interesting patterns nonetheless, Sebastian will have more insights.