This related model-family submission adds 1,689 compressed task results: 563 each for nvidia/Nemotron-3.5-Embed-8B-BF16, nvidia/Nemotron-3.5-Embed-1B-BF16, and nvidia/Nemotron-3.5-Rerank-8B-BF16.
Model
Method
Base Overall nDCG@10
Task files
Variants per task
Nemotron-3.5-Embed-8B-BF16
dense
0.7011
563
15
Nemotron-3.5-Embed-1B-BF16
dense
0.6191
563
15
Nemotron-3.5-Rerank-8B-BF16
reranker
0.7474
563
N/A
Overall follows the current viewer policy: 550 raw tasks grouped into 381 score units; all 563 evaluation tasks, including the 13 overlapping copies, remain in the submitted files. Dense scores above are the base embeddings. The comparison below may select a model's best unquantized truncation variant.
Reranker scores use the dataset-provided reranking_hybrid candidate subsets. They measure candidate reranking and should be interpreted with that candidate protocol when comparing to dense full-corpus retrieval. --rerank-top-k was not overridden.
Command paths below are relative to the code repository root. Replace GPU_ID with physical GPU 0 or 1; cuda:0 refers to the single visible GPU for each worker.
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 component
nvidia/Nemotron-3.5-Embed-8B-BF16 (4096 dims)
nvidia/Nemotron-3.5-Embed-1B-BF16 (2048 dims)
nvidia/Nemotron-3.5-Rerank-8B-BF16
Qwen/Qwen3-Embedding-0.6B (1024 dims)
jinaai/jina-embeddings-v5-text-small (1024 dims)
BAAI/bge-m3 (1024 dims)
intfloat/multilingual-e5-small (384 dims)
bm25
Overall
0.7011
0.6191
0.7474
0.5944
0.6296
0.5820
0.5149
0.4806
Add HAKARI-Bench results for nvidia/Nemotron-3.5-Embed-8B-BF16
BF16, FlashAttention 2, maximum sequence length 32768, no remote code, and pinned model revisions from the linked model cards. SentenceTransformers 6.1.0 with the image extra, Transformers 5.12.1, torch 2.9.0, torchvision 0.24.0, and FlashAttention 2.8.3 were used. The model-defined query/document prompts resolve to query: and passage: . Mean pooling, L2 normalization, truncation to 1024/512, and all default full/truncated int8/binary and rescore variants are preserved (15 entries per task).
All submitted results use batch size 1. An interrupted worker was resumed using existing per-task caches; only complete task results were retained.
Standard --all coverage: all 563 tasks. Every file was decompressed and audited for identity, revision, dataset revision, runtime, candidate ranking, and intended variant coverage. No maximum-length reduction was used.
Checklist
Result files are committed under hakari-results/nvidia__Nemotron-3.5-Embed-8B-BF16/.
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 nvidia/Nemotron-3.5-Embed-1B-BF16
BF16, FlashAttention 2, maximum sequence length 32768, no remote code, and pinned model revisions from the linked model cards. SentenceTransformers 6.1.0 with the image extra, Transformers 5.12.1, torch 2.9.0, torchvision 0.24.0, and FlashAttention 2.8.3 were used. The model-defined query/document prompts resolve to query: and passage: . Mean pooling, L2 normalization, truncation to 1024/512, and all default full/truncated int8/binary and rescore variants are preserved (15 entries per task).
All submitted results use batch size 1. The initial batch-size-4 smoke run failed with ragged image-placeholder inputs; it was retried at batch size 1 without shortening the model maximum length.
Standard --all coverage: all 563 tasks. Every file was decompressed and audited for identity, revision, dataset revision, runtime, candidate ranking, and intended variant coverage. No maximum-length reduction was used.
Checklist
Result files are committed under hakari-results/nvidia__Nemotron-3.5-Embed-1B-BF16/.
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 nvidia/Nemotron-3.5-Rerank-8B-BF16
BF16, FlashAttention 2, maximum sequence length 32768, no remote code, and pinned model revisions from the linked model cards. SentenceTransformers 6.1.0 with the image extra, Transformers 5.12.1, torch 2.9.0, torchvision 0.24.0, and FlashAttention 2.8.3 were used. The standard SentenceTransformers CrossEncoder loader and model-defined scoring head were used without a prompt override.
All submitted results use batch size 1. The primary forward-order worker ran on GPU 0. After dense evaluation completed, GPU 1 evaluated the remaining tail in reverse task order with the same model card, batch size, package environment, and candidate settings. It wrote to a separate root and published only completed, validated results into the canonical root with no-clobber hard links. The workers eventually overlapped on the final task; the first completed result was retained and the duplicate worker was stopped after all coverage audits passed.
Standard --all coverage: all 563 tasks. Every file was decompressed and audited for identity, revision, dataset revision, runtime, candidate ranking, and intended variant coverage. No maximum-length reduction was used.
Checklist
Result files are committed under hakari-results/nvidia__Nemotron-3.5-Rerank-8B-BF16/.
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.
