Add SlideOn and mSlideOn sparse retrieval results

This related two-model submission contains 728 .json.xz task files, with all 17 sparse representations per task. No generated database, cache, HTML report, model YAML, or code is included.

Code, reviewed model cards, loader and reproduction notes: https://github.com/hakari-bench/hakari-bench/pull/15

Canonical base nDCG@10 (0–1): mSlideOn 0.6185 over 381 grouped units from 550 raw tasks; SlideOn 0.5041 over its English/code 152-unit scope. These different scopes are not directly comparable. The common English/code base scores (0–100) are mSlideOn 51.2707 and SlideOn 50.4129.

Add HAKARI-Bench results for lightonai/mSlideOn

Summary

Overall nDCG@10

Reproducibility

Command

CUDA_VISIBLE_DEVICES=0 PYTHONPATH=$PWD OMP_NUM_THREADS=8 \
uv run --no-project --python /path/to/slideon-venv/bin/python \
python -m hakari_bench.cli evaluate from-model-card \
  --model-card config/model_cards/lightonai__mSlideOn.yaml \
  --params-json '{"model":{"loader":"examples.custom_backends.slideon_loader:load_model"}}' \
  --all --device cuda:0 --batch-size 8

Submitter Notes

• Use the reviewed custom loader to preserve the upstream FP32 sparse head with bf16 backbone/projections, official BF16 head autocast and FP16 retrieval-weight rounding. Model code is reviewed and pinned to the revision above; saved query/document prompt names and internal preprocessing are retained.
• Native 8192-token limit, batch size 8, FlashAttention 2, no length reductions. Base is unpruned. All 16 standard query 8/16/24/32 × document 64/128/256/512 max-active-feature variants exist for every task (17 representations in total). These are nonzero feature budgets in a 32768-dimensional latent space, not dense dimensions.
• Task results record their resolved dataset revision and complete environment. Existing complete results were reused when resuming. The post-encode sparse truncation performance fixes preserve selected feature indices/values and evaluation settings; these fixes are merged in GitHub PR #14.
• Complete standard --all coverage: 563 task files. Overall excludes 13 overlapping NanoBEIR copies, retaining 550 raw tasks; multilingual MNanoBEIR languages are averaged into 13 task-family units, giving 381 grouped score units. The reported Overall is grouped, not a flat 550-task mean. All tasks used GPU 0.

Checklist

• Result files are committed under hakari-results/lightonai__mSlideOn/.
• 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.
• Any non-default prompt, sequence length, attention implementation, candidate ranking, or reranker setting is documented above.

Add HAKARI-Bench results for lightonai/SlideOn

Summary

Overall nDCG@10

Reproducibility

Command

CUDA_VISIBLE_DEVICES=0 PYTHONPATH=$PWD OMP_NUM_THREADS=8 \
uv run --no-project --python /path/to/slideon-venv/bin/python \
python -m hakari_bench.cli evaluate from-model-card \
  --model-card config/model_cards/lightonai__SlideOn.yaml \
  --params-json '{"model":{"loader":"examples.custom_backends.slideon_loader:load_model"}}' \
  --all --english --device cuda:0 --batch-size 8

Submitter Notes

• Use the reviewed custom loader to preserve the upstream FP32 sparse head with bf16 backbone/projections, official BF16 head autocast and FP16 retrieval-weight rounding. Model code is reviewed and pinned to the revision above; saved query/document prompt names and internal preprocessing are retained.
• Native 8192-token limit, batch size 8, FlashAttention 2, no length reductions. Base is unpruned. All 16 standard query 8/16/24/32 × document 64/128/256/512 max-active-feature variants exist for every task (17 representations in total). These are nonzero feature budgets in a 32768-dimensional latent space, not dense dimensions.
• Task results record their resolved dataset revision and complete environment. Existing complete results were reused when resuming. The post-encode sparse truncation performance fixes preserve selected feature indices/values and evaluation settings; these fixes are merged in GitHub PR #14.
• Intentional English/code --all --english coverage: 165 task files, including overlapping NanoBEIR copies; the canonical English scope contains 152 score units. This partial English scope must not be labeled a full multilingual Overall.
• Five initial smoke/English tasks used physical GPU 1 before the user's GPU-0-only restriction. All subsequent model inference used physical GPU 0; completed early results were retained.
• cuSPARSE SpGEMM insufficient resources errors affected three NanoDAPFAM tasks: NanoDAPFAMAllTitlAbsClmToTitlAbsClm, NanoDAPFAMInTitlAbsClmToTitlAbsClm, and NanoDAPFAMOutTitlAbsClmToTitlAbsClm. Each was rerun with the same card, dtype, batch size, prompts and sequence limit, selecting --dataset hakari-bench/NanoDAPFAM --split TASK --retrieval-score-device cpu in place of --all --english. Encoding still used GPU 0; only retrieval scoring and embedding postprocessing ran on CPU. Per-task metadata preserves these exceptions. CPU/GPU accumulation may differ slightly numerically.

Checklist

• Result files are committed under hakari-results/lightonai__SlideOn/.
• 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.
• Any non-default prompt, sequence length, attention implementation, candidate ranking, or reranker setting is documented above.