Length-based large language models
A computer that includes a processor and a memory, the memory including instructions executable by the processor to receive a prompt for a large language model, the prompt including an input text and a target length. The large language model can generate an output text that includes a number of words equal to the target length within a user determined tolerance based on a length guidance embedding vector that encodes the target size.
1 . A system, comprising: a computer that includes a processor and a memory, the memory including instructions executable by the processor to:
receive a prompt for a large language model, the prompt including an input text and a target length wherein the large language model is modified by adding a length guidance vector; and
generate, in the large language model, an output text that includes a number of words equal to the target length within a user determined tolerance based on the length guidance vector that encodes the target length.
2 . The system of claim 1 , wherein the input text includes more words than the target length and the output text is based on the input text.
3 . The system of claim 1 , the instructions including further instructions to receive the input text by a tokenizer that generates tokens that represent words in the input text.
4 . The system of claim 1 , wherein the large language model includes an embedding block that includes an array that includes token vectors and a position vector that encodes the position of the token vectors in the array.
5 . The system of claim 1 , wherein a decoder portion of the large language model generates elements of the length guidance vector based on the target length.
6 . The system of claim 5 , wherein the decoder portion of the large language model encodes the target length by determining a scalar multiple starting at zero at an origin of the length guidance vector and ending at one at an entry equal to the target length and sets a remainder of entries in the length guidance vector to zeros.
7 . The system of claim 6 , wherein the decoder portion of the large language model encodes the target length by determining a scalar multiple equal to a sinusoidal function starting at zero at an origin of the length guidance vector, having a value of one an entry equal to one-half the target length, and returning to 0 at the entry equal to the target length and setting a remainder of entries in the length guidance vector to zeros.
8 . The system of claim 7 , wherein the large language model receives as input a first target length and a second target length indicating a range of target lengths and the decoder portion of the large language model is modified to include a first length guidance vector and a second length guidance vector.
9 . The system of claim 8 , wherein the first length guidance vector includes a first sinusoidal function which determines a scalar multiple beginning at an origin of the first length guidance vector at zero, rises to one at one-half the first target length and falls to zero at the first target length and sets a remainder of entries in the first length guidance vector to zeros.
10 . The system of claim 9 , wherein the second length guidance vector includes a second sinusoidal function which determines a scalar multiple beginning at an origin of the second length guidance vector at zero, rises to one at one-half the second target length and falls to zero at the second target length and sets a remainder of entries in the second length guidance vector to zeros.
11 . The system of claim 10 , wherein the first length guidance vector is added to the second length guidance vector.
12 . The system of claim 1 , wherein the user determined tolerance is selected by the user and determined during training of the large language model and is configurable at inference time.
13 . A method, comprising:
receiving a prompt for a large language model, the prompt including an input text and a target length wherein the large language model is modified by adding a length guidance vector; and
generating, in the large language model, an output text that includes a number of words equal to the target length within a user determined tolerance based on the length guidance vector that encodes the target length.
14 . The method of claim 13 , wherein the input text includes more words than the target length and the output text is based on the input text.
15 . The method of claim 13 , further comprising receiving the input text by a tokenizer that generates tokens that represent words in the input text.
16 . The method of claim 13 , wherein the large language model includes an embedding block that includes an array that includes token vectors and a position vector that encodes the position of the token vectors included in the array.
17 . The method of claim 13 , wherein a decoder portion of the large language model generates elements of the length guidance vector based on the target length.
18 . The method of claim 17 , wherein the decoder portion of the large language model encodes the target length by determining a scalar multiple starting at zero at an origin of the length guidance vector and ending at one at an entry equal to the target length and sets a remainder of entries in the length guidance vector to zeros.
19 . The method of claim 18 , wherein the decoder portion of the large language model encodes the target length by determining a scalar multiple equal to a sinusoidal function starting at zero at an origin of the length guidance vector, having a value of one an entry equal to one-half the target length, and returning to zero at the entry equal to the target length and setting a remainder of entries in the length guidance vector to zeros.
20 . The method of claim 13 , wherein the large language model receives as input a first target length and a second target length indicating a range of target lengths and the decoder portion of the large language model is modified to include a first length guidance vector and a second length guidance vector.