Token Probability Plot Calculator

Analysis results

Load data and calculate results.

Waiting for data

Distribution comparison

Two datasets at the same step are required for divergence statistics.

Interpretation

Sampling preview

Calculated token table

Sort by headers, search tokens, and paginate the processed results.

Token	Token ID	Dataset	Step	Raw logit	Adjusted logit	Original probability	Adjusted probability	Log probability	Rank	Cumulative	Surprise	Selected	Filtered

Token data

Input mode

Active dataset

Active generation step

Probability sum tolerance

Paste token data

Headers are optional. CSV and tab-separated values are supported.

Paste format

Token	Token ID	Logit	Probability	Step	Dataset	Selected	Category	Frequency	Remove

Sampling and decoding controls

Temperature

Top-k

Use zero to disable.

Top-p

Minimum probability

Typical-p

Repetition penalty

Frequency penalty

Presence penalty

Logit bias token

Logit bias amount

Sampling method

Sampling seed

Plot configuration

Plot type

Tokens shown

Sort tokens

Sort direction

Probability scale

Value display

Chart height

Label font size

Token label rotation

Bar spacing

Custom chart title

X-axis label

Y-axis label

Low-probability threshold

Custom highlighted token

Display options

Show data labels Show gridlines Show legend

Formula used

Stable softmax: pᵢ = exp((zᵢ − max(z))/T) ÷ Σⱼ exp((zⱼ − max(z))/T)

Entropy: H = −Σᵢ pᵢ ln(pᵢ)

Perplexity: PP = exp(H)

Negative log-likelihood: NLL = −ln(p selected)

KL divergence: D_KL(P ∥ Q) = Σᵢ P(i) ln(P(i) ÷ Q(i))

Jensen–Shannon divergence: JSD(P,Q) = ½D_KL(P ∥ M) + ½D_KL(Q ∥ M)

How to use

Enter tokens with logits or probabilities.
Assign datasets and generation steps for comparisons.
Choose temperature, filtering, penalties, and sampling settings.
Select a plot type and display configuration.
Calculate, inspect metrics, and export the analysis.

Example data

Token	Logit	Approximate probability	Selected
learning	4.20	42%	Yes
model	3.65	24%	No
system	3.10	14%	No
network	2.70	9%	No
data	2.30	6%	No

Frequently asked questions

What is a token probability?

It is the model probability assigned to one candidate token. Higher values indicate stronger preference at that generation step.

What is a logit?

A logit is an unnormalized model score. Softmax converts all candidate logits into probabilities that sum to one.

Why use stable softmax?

Subtracting the largest logit prevents numerical overflow. It preserves the final probability distribution.

How does temperature change probabilities?

Lower temperature sharpens the distribution. Higher temperature spreads probability across more tokens.

What does top-k filtering do?

It keeps only the k highest-probability candidates. Remaining probabilities are normalized again.

What does top-p filtering do?

It keeps the smallest ranked token set reaching a chosen cumulative probability. This adapts candidate count to uncertainty.

What is typical-p filtering?

It favors tokens whose information content stays near distribution entropy. It can remove unusually surprising candidates.

What does entropy measure?

Entropy measures uncertainty across candidate tokens. Lower entropy indicates a more concentrated prediction.

How is perplexity interpreted?

Perplexity represents effective uncertainty. Smaller values mean fewer equally plausible token choices.

Why can the selected token differ from rank one?

Random sampling can select lower-ranked candidates. Filtering and temperature also change selection probabilities.

What are KL and Jensen–Shannon divergence?

They compare probability distributions. Jensen–Shannon divergence is symmetric and usually easier to interpret safely.

Can this calculator compare generation steps?

Yes. Add rows with matching datasets and different step numbers, then choose heatmap or trajectory plots.