Add Helmet data module for HELMET benchmark datasets

[Jantory]

LongBenchV2 was the only supported dataset for long-context fine-tuning. This PR adds a Helmet data module (keys_values/data/helmet.py) that loads any HELMET benchmark dataset via the existing load_helmet_dev_eval function, enabling fine-tuning on a broader set of long-context tasks (RAG, summarization, ICL, synthetic retrieval, etc.).

What changed:

• keys_values/data/helmet.py: New Helmet class inheriting SequenceLengthFilteredDataModule. It loads dev/eval splits from load_helmet_dev_eval, maps the pre-formatted input/output fields directly to SFTDataset (no prompt construction or truncation needed), and uses the dev split for train/val and the eval split as the test set. Sequence lengths are computed at load time for the sampler.
• keys_values/data/init.py: Exports Helmet alongside LongBenchV2.
• keys_values/finetune/longcontext_full.py: Guarded the unconditional data.metadata_dir = str(init_out_dir(...)) call so it is skipped when metadata_dir is None, allowing data modules other than LongBenchV2 to be used without crashing.

Helmet can be passed via the existing --data CLI argument using jsonargparse's class/init-args syntax, requiring no changes to the training entry point.

A use case for specifying the Helmet dataset and max sequence length can be
python keys_values/__main__.py finetune_long_lora \ "${CHECKPOINT_DIR}" \ --out_dir "${OUT_DIR}" \ --devices 2 \ --data Helmet --data.dataset_key json_kv --data.max_length 64k \ --data.max_seq_length 32768 --data.metadata_dir "${METADATA_DIR}" \ --head_model next_token_prediction \ --precision bf16-true --verbose some \ --kv_cache.name h2o-default --kv_cache.cache_length 16384 --kv_cache.chunk_size 1024 \ --train.save_interval 10 --train.micro_batch_size 4 --train.global_batch_size 8 \ --eval.interval 10

---

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.