Feature Extraction
PEFT
ONNX
Safetensors
sparse-retrieval
splare
opensearch
lora
multilingual

Commit History

Yesterday

  1. Card: Data row says each row is used once with each set of 8 hard negatives

    Verified signature
    facee50
  2. Card: training data are 88,165 query rows with 56,195 distinct questions (MIRACL reuses Mr.TyDi's questions in 11 languages)

    Verified signature
    fcc34bf

11 days ago

  1. Drop chat_template.jinja (unused by an encoder); card metadata: pipeline_tag feature-extraction

    Verified signature
    36ae789
  2. RESULTS: GPU column names its endpoint (handler predates inference.py, same vectors); Speed notes the load generator's vCPUs

    Verified signature
    24b0eb3
  3. Card: size against mlv1, after the speed table

    Verified signature
    e1ef020
  4. Card: OpenSearch snippets compact, wrapped only where a line would scroll

    Verified signature
    6ee1b8f
  5. Card: OpenSearch with both routes (own vectors with query_tokens; the model in ml-commons via an ingest pipeline and model_id), indented snippets

    Verified signature
    d1df32c
  6. Card: training table with one setting per row (data, row, objective, fine-tuning, steps, batch, optimizer, length, hardware)

    Verified signature
    cd0f869
  7. Card: training as a table, with LoRA alpha, FLOPS lambda, steps and hardware

    Verified signature
    d1eef04
  8. RESULTS: drop the 36-set comparison with the same-negatives run (model-selection evidence; kept in the code repo's DETAILS.md)

    Verified signature
    81782bd
  9. Card: speed as a table; RESULTS: Speed moved to the end, after the quality sections

    Verified signature
    8154190
  10. NOTICE: attributions only; RESULTS: GPU and CPU rows name the layers and the Docker host, new Speed section (queries and ingest); card: speed summary

    Verified signature
    71348d0
  11. Card: drop Limitations; license sentence links the Qwen LICENSE files at the pinned revisions; absolute NOTICE and RESULTS links

    Verified signature
    50f72ea

12 days ago

  1. Split results by query type; name the Gemma-2 SAE size

    Verified signature
    4af0a86
  2. Initial upload

    Verified signature
    1a8e0c0
  3. initial commit

    Verified signature
    5ce2c63