Add SFT validation eval with val_data

[philippnormann]

Summary

• Add optional val_data and eval config blocks to SFT.
• Run periodic validation inside SFT training and log val/loss and val/num_batches.
• Add config validation that requires eval and val_data to be set together.
• Add unit tests for config validation behavior.

Why

Train loss alone is not enough for checkpoint selection and overfitting detection.

Before

• No native periodic validation signal in SFT runs.

After

• SFT can emit validation metrics at configurable intervals during training.

Evidence

• Reverse-text run showing periodic validation logging behavior.

train/loss	val/loss

• Config used:

sft_fullft_rtext_split_200.toml

max_steps = 200

[ckpt]
interval = 20

[model]
name = "PrimeIntellect/Qwen3-0.6B"

[data]
name = "willcb/R1-reverse-wikipedia-paragraphs-v1-1000"
splits = ["train[:90%]"]
seq_len = 4096
batch_size = 32
shuffle = true
seed = 42

[val_data]
name = "willcb/R1-reverse-wikipedia-paragraphs-v1-1000"
splits = ["train[90%:]"]
seq_len = 4096
batch_size = 32
shuffle = false
seed = 42

[eval]
interval = 10
num_batches = 4

[optim]
lr = 2e-5

Validation

• uv run pytest tests/unit/train/sft/test_sft_eval_config.py -q
• Unit tests cover: eval without val_data (invalid), val_data without eval (invalid), and eval + val_data (valid).
• 200-step reverse-text run emits val/loss every 10 steps as configured.

Scope

• This PR covers periodic SFT validation evaluation and config validation.

---

Note

Medium Risk
Touches core SFT training-loop and dataset-loading paths; while changes are straightforward, they can impact training correctness/performance and distributed metric aggregation if edge cases are missed.

Overview
Adds optional SFT validation via new SFTValConfig (sft.val) that loads a separate validation dataset and runs full-pass evaluation on a configurable interval (and optionally at step 0), logging val/loss.

Refactors SFT data loading by extracting load_sft_dataset() (expensive HF I/O) and extending setup_dataset() to accept a preloaded raw dataset plus max_epochs, enabling validation to reuse preloaded data.

Restructures the SFT training loop to centralize loss/forward-backward into helpers and to compute aggregated loss/NaN counts consistently across distributed ranks, with minor logging tweaks (e.g., only emitting max_vio when present).

^{Written by Cursor Bugbot for commit d80bb23. This will update automatically on new commits. Configure here.}

[samsja]

we don't need a function here

[samsja]

lets not do this, lets rather edit the if statement below to also include step 0 as valid if eval_on_start is set

[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.}

[hallerite]

@philippnormann thank you for the PR! Sorry for taking so long to get it into a mergeable state – it was trickier than we expected, but I think it's in a good state now.

[philippnormann]

No worries, glad it made it in! Appreciate you and @samsja taking the time to get it into shape.

Also have #1849 open for SFT LoRA support if you get a chance to look at it.