rlcli sample for quick generation from base models or checkpoints, and passthrough commands (checkpoint, run, session) that delegate to the official tinker CLI against your local server.
Sampling
Userlcli sample to generate text from a base model or a fine-tuned checkpoint. Sampling uses raw generation with no chat template applied.
Sample from a base model
Sample from a checkpoint
Common options
rlcli uses the model’s tokenizer for encode and decode. You can override the tokenizer model name with the
RLCLI_TOKENIZER environment variable.Managing checkpoints and runs
rlcli provides thin passthrough commands to the official tinker CLI. These setTINKER_BASE_URL to your local server and TINKER_API_KEY=tml-dummy if unset, then invoke python -m tinker.cli <command>.
List checkpoints
Inspect runs
Session management
The passthrough commands are
rlcli checkpoint, rlcli run, and rlcli session. They are not native rlcli implementations but are pre-configured to talk to your local SkyRL server.Next steps
- See
rlcli sampleCLI options - Learn about Tinker Passthrough commands
- Start supervised fine-tuning to create checkpoints you can sample from