Voxiesz commited on
Commit
66f3574
·
verified ·
1 Parent(s): 6b1d663

Sync frozen benchmark results

Browse files
Files changed (2) hide show
  1. README.md +49 -0
  2. 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
+ }