Add SFT dashboard views and workload filtering

[omkaark]

SFT runs now show training loss and supervised targets instead of empty generated responses and reward charts. Adds an SFT/RL filter and labels, a loss chart using the existing chart component, and a Training samples tab that preserves exact target text and labels context separately.

Example run (dashboard here, Qwen3 32B, 8xH200, 1000 steps, 128 bsz, 4096 max ctx len, not tuned):

import hashlib

from pydantic.dataclasses import dataclass
from modal_training_gym import ModelArchitecture, Qwen3_8B_Recipe, SFTDataset
from modal_training_gym.common.models.base import HFModelConfiguration
from modal_training_gym import TrainConfig


class Qwen32B(HFModelConfiguration):
    model_name = "Qwen/Qwen3-32B"
    architecture = ModelArchitecture(
        num_layers=64,
        hidden_size=5120,
        ffn_hidden_size=25600,
        num_attention_heads=64,
        num_query_groups=8,
        kv_channels=128,
        vocab_size=151936,
        untie_embeddings_and_output_weights=True,
        rotary_base=1000000,
    )


@dataclass
class Qwen32BRecipe(Qwen3_8B_Recipe):
    make_vocab_size_divisible_by: int = 16
    optimizer_cpu_offload: bool = False
    pipeline_model_parallel_size: int = 1
    context_parallel_size: int = 1


class FilteredUltraChat(SFTDataset):
    def cache_key(self):
        key = f"{super().cache_key()}:qwen3-32b:max4096:v1"
        return hashlib.sha256(key.encode()).hexdigest()

    def _load_hf_dataset(self):
        from transformers import AutoTokenizer

        tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-32B")
        dataset = super()._load_hf_dataset()
        original_count = len(dataset)

        def fits(batch):
            tokens = tokenizer.apply_chat_template(
                batch["messages"], tokenize=True, return_dict=False
            )
            return [len(row) <= 4096 for row in tokens]

        dataset = dataset.filter(fits, batched=True, batch_size=128)
        print(
            f"UltraChat: kept {len(dataset)} / {original_count} conversations <=4096 tokens",
            flush=True,
        )
        return dataset


def config():
    return TrainConfig(
        model=Qwen32B(),
        dataset=FilteredUltraChat(
            hf_repo="HuggingFaceH4/ultrachat_200k",
            hf_split="train_sft",
            messages_column="messages",
        ),
        recipe=Qwen32BRecipe(
            training_type="sft",
            gpu_type="H200",
            num_gpus_per_node=8,
            tensor_model_parallel_size=8,
            sequence_parallel=True,
            ref_load="/checkpoints/torch_dist/Qwen--Qwen3-32B-tp8-div16-sft-proof",
            num_steps=1000,
            global_batch_size=128,
            max_tokens_per_gpu=4096,
            lr=1e-5,
            save_interval=250,
            shuffle=True,
            max_retries=5,
            image_overlay=lambda image: image.add_local_python_source(
                "sft_ultrachat", copy=True
            ),
            extra_config={"seed": 1234, "rollout_seed": 42},
        ),
    )


if __name__ == "__main__":
    from sft_ultrachat import config

    run = config().launch(prepare_inputs=True)
    print(f"Run ID: {run.training_run_id}")
    print(f"Modal app: {run.modal_app_id}")

Video:
https://github.com/user-attachments/assets/251abe78-4583-4e7c-9a52-2b32494f3fab

[github-actions]

Dashboard preview: https://modal-labs-training-gym--training-gym-previews-preview-r-78da35.modal.run/519/dashboard

[github-actions]

Documentation preview: https://modal-labs-training-gym--training-gym-previews-preview-r-78da35.modal.run/519/docs

[devin-ai-integration]

Note

This report is out of date. Scroll down for Devin Review's latest report on this PR.

Devin Review found 4 new potential issues.

[devin-ai-integration]

🟡 Failed loss requests erase history

An HTTP failure becomes an empty result in fetchRunSteps. The next poll erases loaded loss history and hides the reporting failure.

Suggested change

	if (!res.ok) return [];
	if (!res.ok) throw new Error(await getErrorFromResponse(res));

---

Was this helpful? React with 👍 or 👎 to provide feedback.

[devin-ai-integration]

🟡 Legacy sample links trap navigation

Opening an SFT rollouts deep link makes selectTab push a summary entry. Back restores rollouts, which immediately pushes summary again.

Prompt for agents

In dashboards/frontend/src/pages/TrainingRunDetailPage.svelte, normalize the unsupported rollouts tab for SFT without adding a history entry. Preserve the existing deep-link compatibility while ensuring Back can leave the run page. Consider replacing the current URL entry or making URL parsing workload-aware after the run type is known.

---

Was this helpful? React with 👍 or 👎 to provide feedback.

[devin-ai-integration]

🔍 SFT samples disappeared from this revision

The description promises a Training samples tab with target inspection and export. This revision hides that tab and stops producing its data.

---

Was this helpful? React with 👍 or 👎 to provide feedback.

[devin-ai-integration]

Devin Review found 1 new potential issue.

⚠️ 1 issue in files not directly in the diff

⚠️ RL JSON reward contract breaks

run get --json removes current_reward from every RL payload and emits current_metric instead. Existing consumers read no reward or fail when they require the established key.