Interactive operations
Queue a forward operation
Queue a forward pass without gradients on an interactive training job, then poll the returned future_id until the status is completed, failed, or cancelled.
POST
POST /v1/training/jobs/{job_id}/forward
Computes the loss and per-token log probabilities of a batch without
accumulating gradients. Use it for evaluation or to score tokens before an
update.
The body, response, and result match
POST /v1/training/jobs/{job_id}/forward-backward.
The future type is forward.
A GLiDE (fastino/GLiDE) job takes decisions, and here a row needs no gold
or soft_target. Each result in loss_fn_outputs holds the row’s option
probabilities as probs, as POST /v1/systemone
serves them. Use it to score held-out rows; nothing is trained. See
Train GLiDE.
The numeric token arrays below are illustrative placeholders. Tokenize your
actual text with the tokenizer for the job’s base_model.
Errors
Before queuing the batch, the route also refuses with
400 the batches
forward-backward
refuses for the same reasons. The longest datum forward accepts can differ
from forward-backward’s.
