tinker

name: tinker description: Run LLM post-training on Tinker with CPU-side orchestration and remote GPU execution. Use when preparing or launching SFT, DPO, or PPO-style runs, checkpointing, sampling checkpoints, or resuming long-running jobs through Hermes Research Agent. version: 1.0.0 author: Hermes Research Agent license: MIT metadata: hermes: tags: [tinker, training, sft, dpo, ppo, checkpoints, lora]

Tinker

Tinker is the only compute backend in Hermes Research Agent v1.

Requirements

• TINKER_API_KEY must be set.
• Use tinker_posttrain for lifecycle management instead of ad hoc shell commands when possible.

Supported methods

• sft: dataset rows need prompt and completion
• dpo: dataset rows need prompt, chosen, and rejected
• ppo: dataset rows need prompt, completion, token-level logprobs, and either advantage or reward

Default operating pattern

1. Validate the config with tinker_posttrain(action="validate_config", ...).
2. Start the run with tinker_posttrain(action="start_run", ...).
3. Let research_loop(action="monitor_run", ...) manage long-running polling and resumption.
4. Use tinker_posttrain(action="sample_checkpoint", ...) for quick qualitative checks.
5. Use tinker_posttrain(action="download_checkpoint", ...) only when the project needs local checkpoint artifacts.

Run hygiene

• Always tie the run to an experiment id and a written hypothesis.
• Save checkpoints regularly.
• Record the outcome with research_state(action="record_result", ...).
• If approval mode is active, expect start_run, resume_run, and stop_run to require an approval record.