Add mLateOn results and complete Nemotron-3-Embed-8B-BF16 coverage

Summary

DuckDB Nano-set Comparison

Computed from DuckDB task_results with the same Overall grouping as this PR body. Quantized and rescore variants are excluded; truncate variants are considered, and each model column uses that model's best Overall variant.

Overall nDCG@10

Reproducibility

Command

HAKARI_MLATEON_SHARD=0 HAKARI_MLATEON_MODEL_BATCH_SIZE=32 \
  HAKARI_MLATEON_EXACT_DOC_BATCH_SIZE=128 HAKARI_MLATEON_EXACT_QUERY_BATCH_SIZE=8 \
  CUDA_VISIBLE_DEVICES=0 uv run --group pylate python tmp/run_mlateon_shard.py

HAKARI_MLATEON_SHARD=1 HAKARI_MLATEON_MODEL_BATCH_SIZE=32 \
  HAKARI_MLATEON_EXACT_DOC_BATCH_SIZE=32 HAKARI_MLATEON_EXACT_QUERY_BATCH_SIZE=2 \
  CUDA_VISIBLE_DEVICES=1 uv run --group pylate python tmp/run_mlateon_shard.py

# Resume/retry configuration used for tasks that did not fit the initial batches:
HAKARI_MLATEON_SHARD=<0-or-1> HAKARI_MLATEON_MODEL_BATCH_SIZE=4 \
  HAKARI_MLATEON_EXACT_DOC_BATCH_SIZE=8 HAKARI_MLATEON_EXACT_QUERY_BATCH_SIZE=1 \
  CUDA_VISIBLE_DEVICES=<0-or-1> uv run --group pylate python tmp/run_mlateon_shard.py

Submitter Notes

• The evaluation follows the model card: PyLate exact MaxSim, FP32, 128-dimensional token embeddings, [Q] / [D] prefixes, 8192-token query/document limits, query expansion disabled, and expansion-token attention disabled.
• The 563 standard tasks were split across two RTX 5090 GPUs. Existing task files were cached when retrying. Tasks that exceeded the initial model or exact-scoring batch size were resumed with smaller batches; batch sizing does not change the embeddings or exact MaxSim scores. The result metadata records the effective batch configuration for every task.
• These are complete standard --all results: 563/563 files, including all 550 non-overlapping Overall tasks. The submitted task set has zero missing, extra, or duplicate tasks, and every .json.xz file passed decompression and JSON/score/metadata validation.
• The Overall grouped nDCG@10 is 0.6423. Across the 550 Overall tasks, task-score Pearson correlation is 0.7930 with LiquidAI/LFM2.5-ColBERT-350M, 0.8190 with lightonai/GTE-ModernColBERT-v1, and 0.8139 with lightonai/ColBERT-Zero.

Nemotron-3-Embed-8B-BF16 coverage completion

This PR also adds the six previously missing NanoMTEB-BR results for nvidia/Nemotron-3-Embed-8B-BF16. The remote result set already contains its other 557 standard tasks, including NanoSSRB. Adding these six files produces 563/563 stored tasks and 550/550 Overall tasks, which restores the model's Overall leaderboard row.

CUDA_VISIBLE_DEVICES=0 PYTORCH_ALLOC_CONF=expandable_segments:True \
  uv run --group flash-attn hakari-bench evaluate from-model-card \
    --model-card config/model_cards/nvidia__Nemotron-3-Embed-8B-BF16.yaml \
    --dataset hakari-bench/NanoMTEB-BR \
    --batch-size 2 \
    --device cuda:0

The initial batch-size-4 attempt exhausted GPU memory on BRTaxQAR before writing any result. It was rerun at batch size 2 without changing sequence length, dtype, attention, prompts, or weights. All six files contain the expected base, int8, binary, int8-rescore, and binary-rescore evaluations and passed integrity/metadata validation. A merged-DuckDB check produced 563 base tasks and an Overall row with 550/550 tasks (base mean 68.2748; reranking mean 68.6437).

Checklist

• Result files are staged under the two documented hakari-results/ model paths.
• Result files are compressed .json.xz; no caches, DuckDB files, HTML reports, or local scratch artifacts are included.
• The result JSON records model revision, dataset revision, runtime configuration, and package versions.
• Overall nDCG@10 above was generated from the submitted result files.
• Non-default sequence length, prefixes, batching, attention, and retry choices are documented above.