Wisent Model

Wisent Model owns conversation datasets, preparation and result comparison. Ster owns model training and conversation execution. Stado selects compute, enforces resource limits and retains job outputs; this repository does not rent machines or hold provider credentials.

Supervised training

wisent-model prepare-sft converts the roleplay batches into training examples. Within a Stado job pinned to the source revision, use the prepared examples with Ster:

wisent-model prepare-sft --batches ./data --output ./sft-examples.json
ster tune sft --model Qwen/Qwen3-8B --examples ./sft-examples.json --chat-template off --precision bf16 --device cuda --epochs 3 --learning-rate 0.0002 --output ./sft-adapter.safetensors
ster tune merge --model Qwen/Qwen3-8B --adapter ./sft-adapter.safetensors --output ./sft-model

The prepared prompts already carry the model's turn format, so the example disables an additional chat template. Ster uses unquantized Safetensors and refuses GPTQ training bases.

Conversation reinforcement learning

ster tune grpo trains whole conversations from the merged checkpoint. A second model plays the user, and a reward model trained with ster tune reward scores assistant turns. Supply the actual checkpoint, prompt set and reward artifact to the Stado job. Follow the job with stado job watch JOB and retrieve its outputs with stado results JOB DIRECTORY.

Data generation

wisent-model generate conversation uses declared prompt variants and styles to append generated conversations. wisent-model generate expand expands existing single-turn records. Both call Brama through the workload's Stado model-router configuration, not a direct provider key. They require STADO_MODEL_ROUTER_URL, WISENT_MODEL_MODEL_ROUTER_TOKEN and STADO_WISENT_MODEL_GENERATION_MODEL, and report rejected or unexpanded records with a nonzero exit.

Evaluation

Compare baseline and adapter conversations through ster converse, using the same prompts and generation settings. Preserve the actual conversations and reward measurements rather than treating word lists or output length as a semantic qualification.

See model evaluation and the owning repository contract for commands, inputs and retained-output interpretation.