How to set weights in multi-class classification in xgboost for imbalanced data?

As you say, scale_pos_weight works for two classes (binary classification). weight can be used for three or more classes. The parameter goes into the xgb.DMatrix function and must contain one value for each observation.

Example:

library(xgboost)
data(iris)

# We'll predict Species
label = as.integer(iris$Species)-1
iris$Species = NULL

# Split the data for training and testing (75/25 split)
n = nrow(iris)
train.index = sample(n,floor(0.75*n))

# For example, pick a weight of 1.5 for label "0", 1.0 for the other Species
weights = sapply(label[train.index], function(x) {ifelse(x == 0, 1.5, 1.0)})

# Train the data using weights
xgb.train = xgb.DMatrix(data=as.matrix(iris[train.index,]), label=label[train.index], weight = weights)

A similar question can be found here.