Ant Ling closes the single-machine Agentic RL loop with Ling-3.0-tiny and AReno
Ant Ling has published a practice article describing how it used the local post-training toolkit AReno on a DGX Spark to perform Agentic RL on Ling-3.0-tiny, using tic-tac-toe as a minimal validation task to complete the four-step loop of baseline evaluation, Reward definition, post-training, and re-evaluation in the same environment. The experiment used GSPO for 400 training steps; rollout/rewards_mean rose from about -0.5 to around 0.4, response_len dropped to about 850 tokens, and tool calling and move selection became more stable in the re-evaluation. Ling-3.0-tiny is a hybrid reasoning model with 7.9B total parameters and 1.3B activated parameters; BF16, FP8, and INT4 versions have been open-sourced, and AReno provides a Docker reproduction tutorial.
Ant Ling has published a practice article describing how it used the local post-training toolkit AReno on a DGX Spark to perform Agentic RL on Ling-3.0-tiny, using tic-tac-toe as a minimal validation task to complete the four-step loop of baseline evaluation, Reward definition, post-training, and re-evaluation in the same environment. The experiment used GSPO for 400 training steps; rollout/rewards mean rose from about -0.5 to around 0.4, response len dropped to about 850 tokens, and tool callin
Matched fieldsBody
