1

Suppose that we give a prompt to an LLM model, such as "what is a banana?", the LLM would start writing spitting tokens out, out of a space of tokens, until it manages to complete a textual output that resembles an answer to the question, right?

My question is:

  • Can I, at each step, get a prompt of possible the "next" token, then choose it myself? How?
caveman
  • 163
  • 7

2 Answers2

1

This should be relatively simple. What you would need to do is:

  1. Load the weights into pytorch (via huggingface or from your local machine)
  2. Input your prompt
  3. Get the output probabilities for each token. (You'll probably just want to take a few of the most likely ones)
  4. Then, after selecting your token, you condition on your selected token.

I've looked through some of Torchtune's code for finetuning and it was quite simple. I expect their API for generation to be quite similar. You can probably start from their generation code:

Torchtune

Jayson Meribe
  • 85
  • 3
1

For Azure OpenAI, use max_tokens=1 and eg logprobs=10

https://learn.microsoft.com/en-us/azure/ai-services/openai/reference:

Parameter Type Required? Default Description
prompt string or array Optional <\|endoftext\|> The prompt or prompts to generate completions for, encoded as a string, or array of strings. <\|endoftext\|> is the document separator that the model sees during training, so if a prompt isn't specified the model generates as if from the beginning of a new document.
max_tokens integer Optional 16 The maximum number of tokens to generate in the completion. The token count of your prompt plus max_tokens can't exceed the model's context length. Most models have a context length of 2048 tokens (except for the newest models, which support 4096).
temperature number Optional 1 What sampling temperature to use, between 0 and 2. Higher values mean the model takes more risks. Try 0.9 for more creative applications, and 0 (argmax sampling) for ones with a well-defined answer. We generally recommend altering this or top_p but not both.
top_p number Optional 1 An alternative to sampling with temperature, called nucleus sampling, where the model considers the results of the tokens with top_p probability mass. So 0.1 means only the tokens comprising the top 10% probability mass are considered. We generally recommend altering this or temperature but not both.
logit_bias map Optional null Modify the likelihood of specified tokens appearing in the completion. Accepts a json object that maps tokens (specified by their token ID in the GPT tokenizer) to an associated bias value from -100 to 100. You can use this tokenizer tool (which works for both GPT-2 and GPT-3) to convert text to token IDs. Mathematically, the bias is added to the logits generated by the model prior to sampling. The exact effect varies per model, but values between -1 and 1 should decrease or increase likelihood of selection; values like -100 or 100 should result in a ban or exclusive selection of the relevant token. As an example, you can pass {"50256": -100} to prevent the <|endoftext|> token from being generated.
user string Optional A unique identifier representing your end-user, which can help monitoring and detecting abuse
n integer Optional 1 How many completions to generate for each prompt. Note: Because this parameter generates many completions, it can quickly consume your token quota. Use carefully and ensure that you have reasonable settings for max_tokens and stop.
stream boolean Optional False Whether to stream back partial progress. If set, tokens are sent as data-only server-sent events as they become available, with the stream terminated by a data: [DONE] message.
logprobs integer Optional null Include the log probabilities on the logprobs most likely tokens, as well the chosen tokens. For example, if logprobs is 10, the API will return a list of the 10 most likely tokens. The API will always return the logprob of the sampled token, so there might be up to logprobs+1 elements in the response. This parameter cannot be used with gpt-35-turbo.
suffix string Optional null The suffix that comes after a completion of inserted text.
echo boolean Optional False Echo back the prompt in addition to the completion. This parameter cannot be used with gpt-35-turbo.
stop string or array Optional null Up to four sequences where the API will stop generating further tokens. The returned text won't contain the stop sequence. For GPT-4 Turbo with Vision, up to two sequences are supported.
presence_penalty number Optional 0 Number between -2.0 and 2.0. Positive values penalize new tokens based on whether they appear in the text so far, increasing the model's likelihood to talk about new topics.
frequency_penalty number Optional 0 Number between -2.0 and 2.0. Positive values penalize new tokens based on their existing frequency in the text so far, decreasing the model's likelihood to repeat the same line verbatim.
best_of integer Optional 1 Generates best_of completions server-side and returns the "best" (the one with the lowest log probability per token). Results can't be streamed. When used with n, best_of controls the number of candidate completions and n specifies how many to return – best_of must be greater than n. Note: Because this parameter generates many completions, it can quickly consume your token quota. Use carefully and ensure that you have reasonable settings for max_tokens and stop. This parameter cannot be used with gpt-35-turbo.
Franck Dernoncourt
  • 3,473
  • 2
  • 21
  • 39