Add HAKARI-Bench results for perplexity-ai/pplx-embed-v1-4B

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

# Initial built-in Nano-set run used TEI at max-batch-tokens 8192 and wrote
# 382/382 built-in standard tasks to:
# output/pplx-embed-v1-4B-tei-8k-cosine-all-nano

PYTHONPATH=$PWD uv run hakari-bench evaluate dense \
  --model perplexity-ai/pplx-embed-v1-4B \
  --model-revision 06456497a00540a582918fe8dcd3a5eabb207772 \
  --model-loader examples.custom_backends.tei_embedding:load_model \
  --model-loader-kwargs-json '{"endpoint":"http://127.0.0.1:18081","model":"perplexity-ai/pplx-embed-v1-4B","query_prompt":"","document_prompt":"","query_prompt_name":null,"document_prompt_name":null,"prompts":{"none":""},"similarity_fn_name":"cosine"}' \
  --dtype fp32 \
  --attn-implementation sdpa \
  --retrieval-score-device cpu \
  --batch-size 8 \
  --evaluation-scope standard \
  --embedding-variant truncate:32,64,128,256,512,768,1024,1536,2048 \
  --results-dir output/pplx-embed-v1-4B-tei-8k-cosine-all-nano \
  --result-format json.xz \
  --dataset hakari-bench/NanoBEIR-ar,hakari-bench/NanoBEIR-de,hakari-bench/NanoBEIR-es,hakari-bench/NanoBEIR-fr,hakari-bench/NanoBEIR-it,hakari-bench/NanoBEIR-ja,hakari-bench/NanoBEIR-ko

PYTHONPATH=$PWD uv run hakari-bench evaluate dense \
  --model perplexity-ai/pplx-embed-v1-4B \
  --model-revision 06456497a00540a582918fe8dcd3a5eabb207772 \
  --model-loader examples.custom_backends.tei_embedding:load_model \
  --model-loader-kwargs-json '{"endpoint":"http://127.0.0.1:18082","model":"perplexity-ai/pplx-embed-v1-4B","query_prompt":"","document_prompt":"","query_prompt_name":null,"document_prompt_name":null,"prompts":{"none":""},"similarity_fn_name":"cosine"}' \
  --dtype fp32 \
  --attn-implementation sdpa \
  --retrieval-score-device cpu \
  --batch-size 8 \
  --evaluation-scope standard \
  --embedding-variant truncate:32,64,128,256,512,768,1024,1536,2048 \
  --results-dir output/pplx-embed-v1-4B-tei-8k-cosine-all-nano \
  --result-format json.xz \
  --dataset hakari-bench/NanoBEIR-no,hakari-bench/NanoBEIR-pt,hakari-bench/NanoBEIR-sr,hakari-bench/NanoBEIR-sv,hakari-bench/NanoBEIR-th,hakari-bench/NanoBEIR-vi

# After the second command completed, hakari-bench/NanoBEIR-ko was split to GPU1
# to avoid leaving that GPU idle while GPU0 finished hakari-bench/NanoBEIR-ja.
# GPU0 was interrupted after NanoBEIR-ja reached 13/13 to avoid duplicate ko writes.

Submitter Notes

• Runtime used TEI Docker ghcr.io/huggingface/text-embeddings-inference:120-1.9, model revision 06456497a00540a582918fe8dcd3a5eabb207772, --dtype fp32, --attn-implementation sdpa, --batch-size 8, and cosine similarity. TEI was started with --max-batch-tokens 8192, --max-client-batch-size 8, and --payload-limit 64000000.
• No query/document prompt override was used; query_prompt, document_prompt, query_prompt_name, and document_prompt_name are all null in the result JSON. Dense default quantized/rescore variants plus truncate:32,64,128,256,512,768,1024,1536,2048 were evaluated, producing 50 embedding evaluations per task.
• This PR completes the previously submitted pplx-embed-v1-4B Nano-set coverage by adding the missing MNanoBEIR non-English NanoBEIR slices: ar, de, es, fr, it, ja, ko, no, pt, sr, sv, th, and vi (13 tasks each, 169 files total). Coverage audit after the run: model-card target 551/551 present and MNanoBEIR 182/182 present.
• The model supports 32K context, but TEI evaluation was limited to 8K input tokens because of TEI memory constraints; long-token tasks that require 32K may understate the model's intended long-context performance.

Checklist

• Result files are committed under hakari-results/perplexity-ai__pplx-embed-v1-4B/.
• 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.