launching training

launching from the web ui

the web ui handles environment bundling, dataset upload, and job submission. the final step is the launch screen.

before you launch

review the summary on the launch step. it shows:

• task: your system prompt and completion tags
• dataset: number of training and eval examples
• rewards: configured reward components and their weights
• tools (RAG only): search tool configuration and modes

if anything looks wrong, go back to the relevant step and fix it. you can navigate freely between steps without losing state.

launching

click launch. the web ui will:

1. upload your dataset splits to storage
2. build your environment (generates Python code, runs it remotely, pickles and uploads the result)
3. validate the environment (optional smoke test)
4. create the training run record and submit the job

you land on your training run page once the job is submitted. GPUs take a few minutes to warm up.

downloading as code

to inspect or modify what was generated, click export configuration on the launch step. you can download your config as a .py, .ipynb, or .json file, with an option to include your dataset. run the exported code locally to iterate or customize beyond what the web ui offers.

launching from the python sdk

once you have a dataset and an environment, one call launches the full training pipeline.

quickstart

import trainer
from trainer.corpus.corpora.search import CorporaSearch
from trainer.envs.search_env import SearchEnv
from trainer.trainer.pipeline import train

search = CorporaSearch(
    api_key=API_KEY,
    corpus_name="my-docs",
    base_url=BASE_URL,
)

experiment_id = train(
    env_class=SearchEnv,
    env_args={"search": search},
    train_dataset=train_data,
    eval_dataset=eval_data,
    prefix="my-search",
    api_key=API_KEY,
    local_modules=[cgft],
)

train() returns an experiment ID. view it at https://app.castform.com/experiments/{experiment_id}.

parameters

experiment_id = train(
    env_class=SearchEnv,              # your environment class
    env_args={"search": search},      # constructor args (SearchClient, judge config, etc.)
    train_dataset=list_of_dicts,      # training examples
    eval_dataset=list_of_dicts,       # evaluation examples
    prefix="my-search",            # namespace prefix for upload paths
    api_key=API_KEY,
    local_modules=[cgft],
    dry_run=False,                    # set True to validate without launching
)

dry run

use dry_run=True to validate your environment bundle, dataset upload, and remote smoke test without launching a training job. useful for catching issues before committing to a run.

result = train(
    env_class=SearchEnv,
    env_args={"search": search},
    train_dataset=train_data,
    eval_dataset=eval_data,
    prefix="my-search",
    api_key=API_KEY,
    dry_run=True,
)
# result is a dict with validation info instead of an experiment ID