Sync frozen benchmark results
Browse files- README.md +49 -0
- results.json +52 -0
README.md
ADDED
|
@@ -0,0 +1,49 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: mit
|
| 3 |
+
task_categories:
|
| 4 |
+
- text-retrieval
|
| 5 |
+
- question-answering
|
| 6 |
+
language:
|
| 7 |
+
- en
|
| 8 |
+
tags:
|
| 9 |
+
- agent-memory
|
| 10 |
+
- locomo
|
| 11 |
+
- beir
|
| 12 |
+
- benchmarks
|
| 13 |
+
pretty_name: FluctlightDB Benchmark Results
|
| 14 |
+
size_categories:
|
| 15 |
+
- n<1K
|
| 16 |
+
---
|
| 17 |
+
|
| 18 |
+
# FluctlightDB — Frozen Benchmark Results
|
| 19 |
+
|
| 20 |
+
Official frozen metrics for the FluctlightDB research paper (June 2025).
|
| 21 |
+
|
| 22 |
+
## Files
|
| 23 |
+
|
| 24 |
+
| File | Description |
|
| 25 |
+
|------|-------------|
|
| 26 |
+
| `results.json` | Full benchmark output — LoCoMo, BEIR SciFact, FAMB |
|
| 27 |
+
|
| 28 |
+
## Key numbers
|
| 29 |
+
|
| 30 |
+
- **LoCoMo evidence recall:** 98.1% (1925/1982 gold spans, k=150, hybrid)
|
| 31 |
+
- **BEIR SciFact nDCG@10:** 0.645 (index mode, ties Chroma)
|
| 32 |
+
- **FAMB macro:** 98% index / 97% agent
|
| 33 |
+
|
| 34 |
+
## Reproduce
|
| 35 |
+
|
| 36 |
+
```bash
|
| 37 |
+
git clone https://github.com/voxmastery/FluctlightDB.git
|
| 38 |
+
cd FluctlightDB
|
| 39 |
+
# See benchmarks/README.md and docs/BENCHMARKS.md
|
| 40 |
+
```
|
| 41 |
+
|
| 42 |
+
## Paper
|
| 43 |
+
|
| 44 |
+
- Draft: https://voxmastery.github.io/FluctlightDB/paper/
|
| 45 |
+
- Card: https://huggingface.co/voxmastery/fluctlightdb-paper
|
| 46 |
+
|
| 47 |
+
## Citation
|
| 48 |
+
|
| 49 |
+
Use [CITATION.cff](https://github.com/voxmastery/FluctlightDB/blob/main/CITATION.cff) from the main repository.
|
results.json
ADDED
|
@@ -0,0 +1,52 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"date": "2025-06-22",
|
| 3 |
+
"harness": "FluctlightDB benchmark suite",
|
| 4 |
+
"mode_index": {
|
| 5 |
+
"beir_scifact": {
|
| 6 |
+
"ndcg_at_10": 0.6451,
|
| 7 |
+
"recall_at_10": 0.7833,
|
| 8 |
+
"recall_at_100": 0.925,
|
| 9 |
+
"query_ms": "4-7",
|
| 10 |
+
"chroma_query_ms": "4-7",
|
| 11 |
+
"embedder": "all-MiniLM-L6-v2 ONNX CPU"
|
| 12 |
+
},
|
| 13 |
+
"famb": {
|
| 14 |
+
"macro": 0.98,
|
| 15 |
+
"paraphrase_recall_at_1": 0.92,
|
| 16 |
+
"provenance_top1": 1.0,
|
| 17 |
+
"persistence": 1.0,
|
| 18 |
+
"confusion_ingest": 1.0,
|
| 19 |
+
"determinism": 1.0
|
| 20 |
+
},
|
| 21 |
+
"famb_agent_mode": {
|
| 22 |
+
"macro": 0.97,
|
| 23 |
+
"paraphrase_recall_at_1": 0.83
|
| 24 |
+
},
|
| 25 |
+
"locomo_full": {
|
| 26 |
+
"conversations": 10,
|
| 27 |
+
"memories_ingested": 8695,
|
| 28 |
+
"top_k": 150,
|
| 29 |
+
"mode": "index",
|
| 30 |
+
"rag_mode": "all",
|
| 31 |
+
"mean_evidence_recall": 0.981,
|
| 32 |
+
"evidence_all_in_context": 0.971,
|
| 33 |
+
"evidence_hits": "1925/1982",
|
| 34 |
+
"has_answer_in_context": 0.379,
|
| 35 |
+
"wall_s": 271,
|
| 36 |
+
"cpu_threads": 2,
|
| 37 |
+
"note": "official LoCoMo metrics; hybrid vector+BM25; embed cache warm"
|
| 38 |
+
},
|
| 39 |
+
"longmemeval_s": {
|
| 40 |
+
"status": "deferred",
|
| 41 |
+
"reason": "CPU-heavy per-question ingest (~30s/Q); run later with throttling",
|
| 42 |
+
"dataset": "/tmp/longmemeval/data/longmemeval_s_cleaned.json",
|
| 43 |
+
"questions": 500,
|
| 44 |
+
"pilot_n20": {
|
| 45 |
+
"answer_in_recall_at_8": 0.70,
|
| 46 |
+
"hits": "14/20",
|
| 47 |
+
"sec_per_question": 30.4
|
| 48 |
+
},
|
| 49 |
+
"harness": "benchmarks/run_longmemeval.sh"
|
| 50 |
+
}
|
| 51 |
+
}
|
| 52 |
+
}
|