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

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 \
  --trust-remote-code \
  --model-max-seq-length 8192 \
  --dataset ... \
  --batch-size 8 \
  --retrieval-score-device cpu \
  --embedding-variant truncate:32,64,128,256,512,768,1024,1536,2048 \
  --results-dir output/pplx-embed-v1-4B-tei-8k-cosine-all-nano

# The full standard Nano target was split across two local TEI endpoints:
#   GPU0: http://127.0.0.1:18081
#   GPU1: http://127.0.0.1:18082
# Both endpoints used ghcr.io/huggingface/text-embeddings-inference:120-1.9,
# TEI 1.9.3, fp32, --max-batch-tokens 8192, and --max-client-batch-size 8.

Submitter Notes

• Model-specific settings: TEI custom backend, cosine similarity, fp32, sdpa, trust-remote-code, batch size 8, CPU retrieval scoring, no query/document prompt, and no prompt names. The model README says instruction prompting is not required.
• Caveat: the model supports 32K context, but this run used 8K inference because TEI 32K/24K/16K warmup failed with CUDA out-of-memory on the tested 96 GiB Blackwell GPUs. Long-token tasks that require the full 32K context may therefore understate the model's intended long-context performance.
• TEI precision caveat: fp16 startup failed with Pplx1 is only supported in fp32 precision, so fp32 was used.
• Coverage: standard all-Nano scope, 382/382 result files. The run was resumed and sharded by dataset/split across two GPUs; final audit found no metadata issues, 382 files with cosine base metrics, and 50 embedding evaluations per task.

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.