Sweep the sampling temperature · Ingénierie IA

A decoder turns the logits of the next token into a distribution and draws from it. Temperature changes how flat that distribution is. Top p then cuts off the tail. In this lab you sweep the temperature on fixed logits and measure the entropy, the top 1 probability and the number of tokens that top p keeps.

support.py gives the 10 token vocabulary and the logits from the GPT lesson demo, LOGITS = [3.2, 1.1, 2.8, 0.4, 0.9, 1.5, -0.2, 2.1, 0.7, 0.1]. It also gives the grid TEMPERATURES from 0.1 to 3.0 in steps of 0.1, TARGET_ENTROPY = 2.0 bits and TOP_P = 0.9.

Task

Write five helper functions and answers().

softmax_t(logits, temperature) returns p_i = e^{z_i / T} / \sum_j e^{z_j / T}. It must not overflow for large logits or a small T.
entropy_bits(probs) returns H = -\sum_i p_i \log_2 p_i in bits. A zero probability adds 0.
sweep_temperature(logits, temperatures) returns two float arrays of the length of temperatures: the entropy and the top 1 probability \max_i p_i at each temperature.
temperature_for_entropy(logits, target, lo=0.05, hi=10.0, iters=60) finds the temperature where the entropy equals target, by bisection. Do iters rounds. In each round, let m be the midpoint of lo and hi. If the entropy at m is below target, move lo to m. Else move hi to m. After the last round, return the midpoint of lo and hi.
kept_count(probs, p) sorts the probabilities from high to low and returns the length of the shortest prefix whose cumulative sum is greater than or equal to p. If no prefix reaches p because of rounding, it returns the number of tokens.

Then do the experiment in answers().

Set the temperature to 1.0. Report the entropy as entropy_t1.
Set the temperature to 0.7. Report the top 1 probability as top1_t07.
Change the temperature over TEMPERATURES. Measure the entropy at each one. Report the first grid temperature with an entropy greater than or equal to TARGET_ENTROPY as first_grid_temperature.
Find the exact crossing with temperature_for_entropy(LOGITS, TARGET_ENTROPY). Report it as crossing_temperature.
Apply top p with TOP_P at the temperatures 0.7, 1.0 and 1.5. Report the kept counts as kept_p09_t07, kept_p09_t10 and kept_p09_t15.

Input and output

Name	Type	Shape	Meaning
`logits`	float array	`(V,)`	one score for each token
`temperature`	float	scalar	the divisor T of the logits, T > 0
`probs`	float array	`(V,)`	a distribution that sums to 1
`temperatures`	list of floats	`(k,)`	the temperatures of the sweep
`target`	float	scalar	the entropy to reach, in bits
`p`	float	scalar	the top p mass, `0 < p <= 1`

softmax_t returns a float array of shape (V,).
entropy_bits returns a float.
sweep_temperature returns a tuple of two float arrays of shape (k,).
temperature_for_entropy returns a float. kept_count returns an int.
answers() returns a dict with the seven keys above. Each value is a Python float.

Constraints

2 <= V <= 1000. The logits are finite and can be larger than 700.
Do not change the input arrays.
Do not round the answers. The checks accept a relative difference of 1e-6 from the reference run.
Use numpy and the standard library only.

Examples

python

softmax_t(np.array([1.0, 2.0, 3.0]), 0.5)

text

[0.01587624 0.11731043 0.86681333]

python

entropy_bits(np.array([0.5, 0.25, 0.25]))

text

1.5

python

sweep_temperature(LOGITS, [0.5, 1.0, 2.0])

text

(array([1.50271532, 2.52573217, 3.09650953]), array([0.61269561, 0.38024152, 0.22636197]))

python

kept_count(np.array([0.1, 0.6, 0.3]), 0.8)

text