Add mLateOn results and complete Nemotron 8B coverage

#33
HAKARI-Bench org

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

Summary

Field Value
Model lightonai/mLateOn
Result directory lightonai__mLateOn
Target path hakari-results/lightonai__mLateOn
Result files 563 total, 563 .json.xz
Evaluation method late-interaction
Overall nDCG@10 0.6423
Overall score units 381 grouped units from 550 raw task results

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 component lightonai/mLateOn LiquidAI/LFM2.5-ColBERT-350M (128 dims) lightonai/GTE-ModernColBERT-v1 (128 dims) lightonai/ColBERT-Zero (128 dims)
Overall 0.6423 0.4943 0.4628 0.4623
NanoMMTEB-v2 0.5371 0.4490 0.4075 0.4250
NanoRTEB 0.6417 0.5208 0.5079 0.5101
MNanoBEIR 0.6333 0.5445 0.4677 0.4657
NanoBIRCO 0.3364 0.2551 0.2335 0.2315
NanoMLDR 0.8274 0.4089 0.3163 0.3095
NanoLongEmbed 0.8009 0.5583 0.5159 0.5514
NanoDAPFAM 0.2835 0.2706 0.2555 0.2556
NanoCoIR 0.8547 0.6284 0.6371 0.6280
NanoIFIR 0.3301 0.2596 0.2338 0.2321
NanoLaw 0.7255 0.4631 0.4456 0.4302
NanoMedical 0.5950 0.5287 0.4829 0.4503
NanoRARb 0.2615 0.2002 0.1657 0.1856
NanoBRIGHT 0.3437 0.3160 0.3043 0.3151
NanoCodeRAG 0.8765 0.8543 0.8420 0.8630
NanoChemTEB 0.8373 0.8710 0.8134 0.8003
NanoR2MED 0.3063 0.2148 0.1669 0.1710
NanoBuiltBench 0.4973 0.5073 0.4878 0.5026
NanoCMTEB 0.7775 0.6763 0.6501 0.6467
NanoIndicQA 0.8032 0.1898 0.2088 0.2411
NanoMuPLeR 0.9764 0.8053 0.7939 0.6804
NanoMTEB-v2 0.6270 0.6103 0.5953 0.6086
NanoMTEB-Dutch 0.6245 0.5257 0.5103 0.5162
NanoMTEB-French 0.6473 0.5942 0.4893 0.4866
NanoMTEB-German 0.7119 0.6214 0.5562 0.5547
NanoJMTEB-v2 0.8502 0.7473 0.6445 0.6352
NanoMTEB-Korean 0.8418 0.8090 0.5545 0.5238
NanoFaMTEB-v2 0.7068 0.5005 0.4848 0.4871
NanoMTEB-Polish 0.5412 0.3840 0.3910 0.3959
NanoMTEB-BR 0.7168 0.6352 0.5452 0.5448
NanoSSRB 0.3723 0.3210 0.3316 0.2877
NanoRuMTEB 0.9117 0.8383 0.8008 0.7632
NanoMTEB-Scandinavian 0.7977 0.7059 0.6416 0.6499
NanoMTEB-Spanish 0.6372 0.5697 0.4347 0.4282
NanoMTEB-Thai 0.7896 0.1724 0.3338 0.3330
NanoVNMTEB 0.6041 0.3974 0.4198 0.4221
NanoMTEB-Misc 0.7878 0.6673 0.5862 0.5428
NanoMIRACL 0.8296 0.6540 0.5890 0.6520

Overall nDCG@10

Overall component nDCG@10 Score units Raw task results
NanoMMTEB-v2 0.5371 18 18
NanoRTEB 0.6417 14 14
MNanoBEIR 0.6333 13 182
NanoBIRCO 0.3364 5 5
NanoMLDR 0.8274 13 13
NanoLongEmbed 0.8009 6 6
NanoDAPFAM 0.2835 12 12
NanoCoIR 0.8547 10 10
NanoIFIR 0.3301 4 4
NanoLaw 0.7255 4 4
NanoMedical 0.5950 7 7
NanoRARb 0.2615 14 14
NanoBRIGHT 0.3437 20 20
NanoCodeRAG 0.8765 4 4
NanoChemTEB 0.8373 3 3
NanoR2MED 0.3063 8 8
NanoBuiltBench 0.4973 2 2
NanoCMTEB 0.7775 8 8
NanoIndicQA 0.8032 11 11
NanoMuPLeR 0.9764 14 14
NanoMTEB-v2 0.6270 10 10
NanoMTEB-Dutch 0.6245 27 27
NanoMTEB-French 0.6473 8 8
NanoMTEB-German 0.7119 5 5
NanoJMTEB-v2 0.8502 11 11
NanoMTEB-Korean 0.8418 5 5
NanoFaMTEB-v2 0.7068 17 17
NanoMTEB-Polish 0.5412 14 14
NanoMTEB-BR 0.7168 6 6
NanoSSRB 0.3723 6 6
NanoRuMTEB 0.9117 3 3
NanoMTEB-Scandinavian 0.7977 7 7
NanoMTEB-Spanish 0.6372 7 7
NanoMTEB-Thai 0.7896 9 9
NanoVNMTEB 0.6041 26 26
NanoMTEB-Misc 0.7878 12 12
NanoMIRACL 0.8296 18 18

Reproducibility

Field Value
Model source lightonai/mLateOn
Model revision 35391e36392085d72a93d232f6122607a234ad7a
Dataset revision(s) 00541a0fce4048057fb7ddec30d37155a5c23d95, 01736efbaa96f020c2a4d996efdacc18071e2fcb, 017849a95097eea984680cbab35972f8d3812376, 0f3a6f43b8a26a9b8c8d5f31b09bd60dc4cd572d, 1726763179e1e114ad9ffcdc7262923471e8ecc8, ... (50 total)
Evaluated at UTC 2026-08-08T09:47:53.523016+00:00 to 2026-08-08T16:21:29.135214+00:00
Generated at UTC 2026-08-08T09:47:53.734064+00:00 to 2026-08-08T16:21:29.135238+00:00
dtype fp32
device cuda:0
batch size 16, 32, 4
attention implementation not recorded
trust remote code False
max sequence length 8191
candidate ranking reranking_hybrid
rerank top-k not recorded
query prompt name not recorded
document prompt name not recorded
Python 3.12.12 (main, Dec 9 2025, 19:02:36) [Clang 21.1.4 ]
Platform Linux-6.8.0-107-generic-x86_64-with-glibc2.39
torch 2.9.0
transformers 5.12.1
sentence-transformers 5.4.1
datasets 4.8.4
CUDA available=True, version=12.8
CUDA devices 0: NVIDIA GeForce RTX 5090

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.

Field Value
Model nvidia/Nemotron-3-Embed-8B-BF16
Target path hakari-results/nvidia__Nemotron-3-Embed-8B-BF16/hakari-bench__NanoMTEB-BR
Added files 6 .json.xz
NanoMTEB-BR mean nDCG@10 0.7203
Model revision 2b29550c4ab0646bb6bb47032dda54ea11f6dfe2
Runtime BF16, Flash Attention 2, 32768-token maximum, 4096 dimensions
Prompts query name query (query: ); document name document (passage: )
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.
hotchpotch changed pull request status to merged

Sign up or log in to comment