Replace results for official Bekko a8m and a25m releases
#27
by hotchpotch - opened
Replace HAKARI-Bench results for the official Bekko a8m and a25m releases
Summary
This PR replaces the existing HAKARI-Bench result directories for the two official Bekko v1 release models with the prefix-free revisions published on 2026-07-19 JST.
| Model | Target path | Files replaced | Revision | Overall nDCG@10 |
|---|---|---|---|---|
hotchpotch/bekko-embedding-v1-a8m |
hakari-results/hotchpotch__bekko-embedding-v1-a8m |
551 .json.xz |
953408def97ea884d8f167a64be5d613d6d2a125 |
0.5453 |
hotchpotch/bekko-embedding-v1-a25m |
hakari-results/hotchpotch__bekko-embedding-v1-a25m |
551 .json.xz |
8acf4b7403ae799684467456c6102529d023e2f2 |
0.5700 |
The old files at both paths are intentionally replaced rather than retained as parallel model entries. The -pt pretraining checkpoints are not part of this PR.
Release Change
- Removes the former
query:/passage:prompt behavior. Both query and document prompts are empty strings in every submitted result. - Pins the new official Hugging Face model revisions listed above.
- Evaluates the native 384-dimensional embedding and MRL truncation at 256, 128, and 64 dimensions.
- Includes the standard float, int8, binary, int8-rescore, and binary-rescore combinations: 20 embedding evaluations in every task file.
- Covers the complete restored
--allscope: 551 task files per model.
Evaluation Settings
| Field | Value |
|---|---|
| Evaluation method | SentenceTransformers dense retrieval |
| Runtime | Python 3.12.11, torch 2.9.0, transformers 5.12.1, sentence-transformers 5.4.1, datasets 4.8.4 |
| Device | NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition |
| CUDA / cuDNN | 12.8 / 91002 |
| dtype | bf16 |
| Attention | flash_attention_2 (FlashAttention 2.8.3) |
| Batch size | 128 |
| Maximum sequence length | 8192 |
| Query / document prompts | empty / empty |
| Trust remote code | false |
| Similarity | cosine |
| Candidate ranking | reranking_hybrid |
| Retrieval score device | auto |
Representative command shape for each model:
uv run hakari-bench evaluate dense \
--model MODEL_ID \
--model-revision MODEL_REVISION \
--all \
--dtype bf16 \
--attn-implementation flash_attention_2 \
--device cuda:0 \
--batch-size 128 \
--model-max-seq-length 8192 \
--embedding-variant truncate:256,128,64 \
--results-dir output/bekko-release-20260719-tf5.12.1-fa2.8.3
No prompt override or prompt name was supplied.
Selected Scores
Scores use the base 384-dimensional embedding, with HAKARI Overall aggregated over 538 deduplicated tasks.
| Model | Overall | MNanoBEIR | NanoMMTEB-v2 | NanoRTEB | NanoCoIR | NanoMLDR | NanoLongEmbed |
|---|---|---|---|---|---|---|---|
| a8m | 0.5453 | 0.5265 | 0.5025 | 0.5498 | 0.7468 | 0.5448 | 0.6819 |
| a25m | 0.5700 | 0.5490 | 0.4937 | 0.5936 | 0.7858 | 0.5714 | 0.7060 |
Validation
- Confirmed 551/551 task files for each model, 1102 files total.
- Confirmed every submitted file is compressed
.json.xz;xz -tpassed for all 1102 files. - Confirmed every file records the exact model revision, empty query/document prompts,
max_seq_length=8192,dtype=bf16, andflash_attention_2. - Confirmed each file contains the intended 20 embedding variants.
- Rebuilt the latest remote leaderboard DuckDB with these two paths replaced and verified the models in the local viewer.
- No DuckDB, cache, YAML, report, or scratch file is included in the PR.
hotchpotch changed pull request status to merged