Add HAKARI-Bench results for litillabs/litil-colbert-300m-v1

Summary

The leaderboard's Overall NanoLaw component counts four of the eight tasks (NanoGerDaLIRSmall, NanoLeCaRDv2, NanoLegalBenchConsumerContractsQA, NanoLegalQuAD) per config/viewer/overall_tasks.yaml; all eight are submitted.

Per-Task nDCG@10 (all 8 submitted NanoLaw tasks)

Values are the late_interaction_exact_maxsim_ndcg@10 entries from each task's metrics key (exact MaxSim headline). The reranking_hybrid diagnostic pass is recorded under rerank_metrics in the result files.

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-bench evaluate from-model-card \
  --model-card config/model_cards/litillabs__litil-colbert-300m-v1.yaml \
  --dataset hakari-bench/NanoLaw \
  --dataset-revision 47dce6a36e9ae44fab47d735c5dec8c16c946495 \
  --evaluation-scope standard \
  --dtype fp32 \
  --device cuda:0 \
  --batch-size 1 \
  --late-interaction-exact-doc-batch-size 16 \
  --late-interaction-exact-query-batch-size 1 \
  --candidate-ranking reranking_hybrid \
  --results-dir output/hakari-results \
  --split NanoAILACasedocs --split NanoAILAStatutes --split NanoGerDaLIRSmall \
  --split NanoLeCaRDv2 --split NanoLegalBenchConsumerContractsQA \
  --split NanoLegalBenchCorporateLobbying --split NanoLegalQuAD \
  --split NanoLegalSummarization

Run on a Hugging Face Jobs a100-large instance (NVIDIA A100-SXM4-80GB, CUDA 12.8); single pass.

Submitter Notes

• Full-parameter fine-tune of lightonai/mLateOn
  (35391e36392085d72a93d232f6122607a234ad7a) for legal retrieval; model card
  PR: https://github.com/hakari-bench/hakari-bench/pull/3.
• Non-default setting: late_interaction.query_length is 1024 (base: 8192),
  as declared in the checkpoint configuration and the model card; chosen on
  the developer's own development data. document_length 8192, fp32,
  trust_remote_code false, no prompt names; only the [Q] / [D]
  prefixes.
• Coverage is intentionally NanoLaw-only: all eight hakari-bench/NanoLaw
  tasks, no other collections. The model targets legal retrieval.
• Single pass, no retries, no batch-size changes, no failed tasks.

Checklist

• Result files are under hakari-results/litillabs__litil-colbert-300m-v1/hakari-bench__NanoLaw/, compressed .json.xz, nothing else.
• Result JSON records the public model revision, dataset revision, runtime configuration and package versions.
• Non-default settings (query length 1024, candidate ranking) are documented above and in the model card.
• Coverage is intentionally partial (NanoLaw only) and stated above.
• Model card PR linked: https://github.com/hakari-bench/hakari-bench/pull/3