Add HAKARI-Bench results for LiquidAI LFM2.5 models

This PR adds all configured HAKARI-Bench task/collection result JSON files for two LiquidAI LFM2.5 models.

Summary

Only .json.xz result files are included. Local sidecar YAML, DuckDB files, reports, and caches are intentionally excluded. The result set spans 48 dataset directories per model, including NanoBEIR language collections and other Nano benchmark collections.

Evaluation settings

LiquidAI/LFM2.5-ColBERT-350M

• Backend: PyLate ColBERT exact MaxSim
• Runtime: Python 3.12.12, torch 2.9.0, transformers 5.3.0, sentence-transformers 5.4.1, datasets 4.8.4
• Device: NVIDIA GeForce RTX 5090, CUDA 12.8
• Dtype: fp32
• Attention: sdpa
• Batch size: 128
• Query/document lengths: 48 / 512
• Query/document prefixes: [Q] / [D]
• Candidate ranking: reranking_hybrid
• Trust remote code: enabled

Representative command:

uv run hakari-bench evaluate late-interaction \
  --model LiquidAI/LFM2.5-ColBERT-350M \
  --model-revision 59633c2e31717b3502343ff566bee9fda3261943 \
  --all \
  --dtype fp32 \
  --attn-implementation sdpa \
  --device cuda:0 \
  --trust-remote-code \
  --batch-size 128 \
  --late-interaction-query-length 48 \
  --late-interaction-document-length 512 \
  --late-interaction-query-prefix "[Q] " \
  --late-interaction-document-prefix "[D] " \
  --late-interaction-exact-doc-batch-size 128 \
  --late-interaction-exact-query-batch-size 8

LiquidAI/LFM2.5-Embedding-350M

• Backend: SentenceTransformers dense embedding
• Runtime: Python 3.12.12, torch 2.9.0, transformers 5.3.0, sentence-transformers 5.4.1, datasets 4.8.4
• Device: NVIDIA GeForce RTX 5090, CUDA 12.8
• Dtype: bf16
• Attention: flash_attention_2
• Batch size: 128
• Prompt names: query / document
• Encode kwargs: {"normalize_embeddings": true}
• Candidate ranking: reranking_hybrid
• Trust remote code: enabled

Representative command:

uv run --group tf5-fa2 hakari-bench evaluate dense \
  --model LiquidAI/LFM2.5-Embedding-350M \
  --model-revision 86b3e1bddcceb99f86104dc996a4fb6169e28f3a \
  --all \
  --dtype bf16 \
  --attn-implementation flash_attention_2 \
  --device cuda:0 \
  --trust-remote-code \
  --batch-size 128 \
  --query-prompt-name query \
  --document-prompt-name document \
  --encode-kwargs-json '{"normalize_embeddings": true}'

Validation

• Local result coverage audit confirmed 551 .json.xz files per model, 1102 total.
• Both model result directories were appended to the latest local DuckDB cache and viewed in the leaderboard viewer.
• The generated model cards and language metadata support were added in the related code change: 4622602 Add LFM model cards and language inference.