train-complex-blackbox
Installation
SKILL.md
0. Ask user for API key + model (or API key + base url + model) for debugging
This is not 100% necessary, but it can help a lot in debugging in step 1. If user has not given a API, ask user to give your one.
By default, the code you write should be located at ./tutorial/opencode_build_xxxxxx/*.py
1. Initial Programming
Writing dataset collector (get_training_dataset_item_list.py)
get_training_dataset_item_list.py: Returns a list of training data items. Maybe a list of training tasks, each item is a string identifier of a training task, or a dict containing necessary information for the training task.
Episode Runner (run_episode_once.py)
-
run_episode_once.py:- Argument Parser: takes (training data item identifier + api-key + base-url) as input, model-name is not required, you can make up a model name because we ignore it.
Related skills
More from modelscope/agentjet
openjudge
>
2write-swarm-client
Create a trainable agent loop or agent workflow with AgentJet
2write-passive-swarm-client
Create a passive swarm client that waits for user input instead of iterating through a dataset by itself.
2map-verl-config
map verl config to agentjet config
1monitor-with-tmux
Monitor training progress by reading tmux content with exponential backoff intervals (30s, 1min, 2min, 4min, 8min, 16min), analyze logs when anomalies occur, and provide fix suggestions
1install-agentjet-client
Install agentjet swarm server with uv package manager
1