Add HAKARI-Bench results for OpenAI text-embedding-3 models

This PR adds .json.xz benchmark result files for two OpenAI embedding models under hakari-results/. It intentionally excludes local model-card YAML, DuckDB files, caches, HTML reports, batch workspaces, and scratch artifacts.

Branch: add-openai-text-embedding-3-results
Commit: 191ebb2e51a2c2b7cc7bde31ac409478e15fec68

Submitted models

Validation

• 1102 staged and pushed files total: 551 .json.xz files per model.
• Non-.json.xz files under the submitted result directories: 0.
• DuckDB / cache / batch workspace artifacts under the submitted result directories: 0.
• Git LFS/Xet upload completed for all 1102 .json.xz files.
• JSON model.id checked across all files: openai/text-embedding-3-small and openai/text-embedding-3-large.
• Dense method checked in JSON.
• Embedding evaluation counts checked in JSON: 20 variants for small and 25 variants for large, across all 551 files each.

Reconstructed commands

The run used task-level OpenAI Batch API registration and processing. The exact internal batch IDs are recorded per result JSON under model.backend_metadata. The practical workflow was:

uv run --group openai hakari-bench batch dense register \
  --provider openai \
  --model text-embedding-3-small \
  --all \
  --results-dir output/hakari-results \
  --max-input-tokens 8100

uv run --group openai hakari-bench batch dense process \
  --provider openai \
  --model text-embedding-3-small \
  --results-dir output/hakari-results

uv run --group openai hakari-bench batch dense register \
  --provider openai \
  --model text-embedding-3-large \
  --all \
  --results-dir output/hakari-results \
  --max-input-tokens 8100

uv run --group openai hakari-bench batch dense process \
  --provider openai \
  --model text-embedding-3-large \
  --results-dir output/hakari-results

Variant policy

• Base full-dimension embeddings are evaluated directly.
• Truncate variants are computed from batch full-dimension embeddings as full[:dim] + L2 normalize.
• Quantized variants include full-dimension int8, binary, int8_rescore, and binary_rescore.
• Truncate variants are also crossed with int8, binary, int8_rescore, and binary_rescore.

Per-model details

Add HAKARI-Bench results for openai/text-embedding-3-small

Summary

Overall nDCG@10

Reproducibility

Reconstructed workflow

uv run --group openai hakari-bench batch dense register \
  --provider openai \
  --model text-embedding-3-small \
  --all \
  --results-dir output/hakari-results \
  --max-input-tokens 8100

uv run --group openai hakari-bench batch dense process \
  --provider openai \
  --model text-embedding-3-small \
  --results-dir output/hakari-results

Submitter Notes

• OpenAI embeddings were generated through the OpenAI Batch API and then materialized into standard HAKARI-Bench per-task .json.xz result files.
• Input texts were truncated to approximately 8100 tokens to stay below the OpenAI embedding input limit with margin.
• Truncation-dimension variants were generated after batch output materialization by taking the full embedding prefix and applying L2 normalization (full[:dim] + normalize). This is close to, but not byte-identical to, OpenAI API-side dimensions output.
• No query or document prompts were used. Retrieval uses cosine similarity and candidate_ranking=reranking_hybrid.
• Parameter counts are unknown for these hosted OpenAI models and are recorded as null/unknown.
• These are full benchmark coverage results for the current 551 submitted result files per model.

Checklist

• Result files are committed under hakari-results/openai__text-embedding-3-small/.
• 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 openai/text-embedding-3-large

Summary

Overall nDCG@10

Reproducibility

Reconstructed workflow

uv run --group openai hakari-bench batch dense register \
  --provider openai \
  --model text-embedding-3-large \
  --all \
  --results-dir output/hakari-results \
  --max-input-tokens 8100

uv run --group openai hakari-bench batch dense process \
  --provider openai \
  --model text-embedding-3-large \
  --results-dir output/hakari-results

Submitter Notes

• OpenAI embeddings were generated through the OpenAI Batch API and then materialized into standard HAKARI-Bench per-task .json.xz result files.
• Input texts were truncated to approximately 8100 tokens to stay below the OpenAI embedding input limit with margin.
• Truncation-dimension variants were generated after batch output materialization by taking the full embedding prefix and applying L2 normalization (full[:dim] + normalize). This is close to, but not byte-identical to, OpenAI API-side dimensions output.
• No query or document prompts were used. Retrieval uses cosine similarity and candidate_ranking=reranking_hybrid.
• Parameter counts are unknown for these hosted OpenAI models and are recorded as null/unknown.
• These are full benchmark coverage results for the current 551 submitted result files per model.

Checklist

• Result files are committed under hakari-results/openai__text-embedding-3-large/.
• 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.