Group-level metadata/metrics/logs should survive round-trip.

Verify group-level metrics are aggregated once per group.

Verify multiple log calls append to history.jsonl.

Verify history.jsonl has correct format for downstream readers.

Verify history.jsonl is readable by pl.read_ndjson (used by delete_checkpoints).

Test handling of invalid JSON.

Non-numeric step suffix should fail with a helpful message.

Test iterate_file reads trajectories correctly.

Test that iterate_file is deterministic with same seed.

Test iterate_file with multiple epochs.

Test iterate_file with initial_skip for resuming.

Tokens with higher KL (more drift) should get reduced advantages.

When ref_logprobs is None, kl_penalty_coef should have no effect.

When kl_penalty_coef>0 and ref_logprobs provided, kl_policy_ref should be set.

When kl_penalty_coef=0, advantages should not be modified.

The KL penalty should be zero-mean across assistant tokens.

Trainable model with step should have hosted_vllm/ prefix and @step suffix.

With step, should append @step suffix to model name.

Without step, should use latest checkpoint name.

Test multi-checkpoint inference with LocalBackend (UnslothService).

Test that long repeated content (like system prompts) compresses well.

LoRA module name should include @step suffix.

LoRA module name should work with step 0.

Test that cosine and linear decay work correctly.

Test LR schedule edge cases.

Test schedule with warmup_steps=0.

Test that warmup doesn't start at lr=0.

max_loras should be configurable via engine_args.

UnslothService should default to max_loras=2 (one for training, one for validation).

Verify metrics are prefixed with split name.

Test migrating an entire directory.

Test dry run mode doesn't modify files.

Test that running migration twice is safe.

Test that delete_original=False preserves the JSONL file.

Test migrating a model directory.

Test migrating a simple JSONL file.

Unsuffixed model names should fail with a helpful message.

Explicitly passing step=None should behave same as no step.

With step, should append :step{N} suffix.

Without step, should return base W&B artifact name.

Test that multiple trajectory groups are preserved separately.

Test multiple trajectories within a single group.

Test that structures without logprobs remain unchanged.

Test that non-JSONL files are skipped.

Test handling of nonexistent file.

Test trajectories with null/empty fields.

Direct parquet reader compatibility.

Verify parquet schema contains expected fields.

Valid `model@step` names should parse correctly.

Verify the API supports the pipelined training pattern from RFC #513. The pattern is: 1. Rollout uses latest checkpoint: mod

Train model A, push to S3, fork from S3 into model B, then train model B.

Train a model, then push its checkpoints to S3.

Test migrating a real trajectory file.

Verify report_metrics can be customized.

Verify report_metrics defaults to None.