feat: add teacher config/sft loss for hosted training SFT

[eligotts]

Summary

Implements APR-157 SFT distillation support through the existing prime train config path.

• Adds top-level loss = "rl" | "sft" with [teacher] and [teacher.sampling].
• Builds teacher sampling on a general sampling-args config so provider/API-specific fields can pass through without adding new CLI schema for every parameter.
• Validates that SFT requires a teacher, RL rejects teacher config, and teacher.save = true is not supported.
• Defaults omitted SFT rollouts_per_example to 1 while preserving explicit overrides.
• Forwards the public platform payload shape as loss and teacher.
• Routes top-level enable_thinking and reasoning_effort sampling aliases into extra_body.chat_template_kwargs.
• Updates the confirmation summary so Training, Teacher, and Run Config render as separate sections.
• Keeps the generated template TOML-safe when SFT teacher config and checkpoint_id are uncommented together.

Config Shape

model = "openai/gpt-oss-20b"
loss = "sft"

[teacher]
model = "openai/gpt-oss-120b"
save = false

[teacher.sampling]
max_tokens = 2048
enable_thinking = false
reasoning_effort = "medium"

[[env]]
id = "primeintellect/wordle"

API Payload

{
  "loss": "sft",
  "teacher": {
    "model": "openai/gpt-oss-120b",
    "save": false,
    "sampling": {
      "max_tokens": 2048,
      "extra_body": {
        "chat_template_kwargs": {
          "enable_thinking": false,
          "reasoning_effort": "medium"
        }
      }
    }
  }
}

Tests

• uv run pytest packages/prime/tests/test_rl_config.py packages/prime/tests/test_train_cli.py -q

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit d2f3992. Configure here.}

[cursor]

Inconsistent truthiness check for teacher reasoning_effort display

Low Severity

The teacher display section checks teacher_sampling.reasoning_effort using bare truthiness (lines 936 and 944), while the equivalent main sampling display section uses is not None checks (lines 964 and 986). Since reasoning_effort is str | None, the truthiness check would suppress display for any falsy string value (e.g. ""), creating inconsistent behavior between the two sections.

Additional Locations (1)

• packages/prime/src/prime_cli/commands/rl.py#L943-L944

 

^{Reviewed by Cursor Bugbot for commit d2f3992. Configure here.}