Ultra-sparse variant of FLES-1. Only 139 non-zero terms per document at threshold=0.3 — comparable to SPLADE's 125. 2.6x sparser than v14 with only 4% lower quality.
This model proved that step-interval CLFR works: a single lambda adjustment after ~6,250 steps outperformed 12,500 per-step adjustments across 5 controller designs.
Metrics
nfcorpus (threshold=0.3)
Metric
Value
NDCG@10
0.2923
MRR
0.5001
Recall@100
0.2367
Avg NNZ
139
Comparison
Model
NDCG@10
NNZ
Index size (1K docs)
fles1-v12b
0.292
139
0.12 MB
fles1-v14
0.305
359
0.32 MB
BM25 (Pyserini)
0.325
—
—
SPLADE-Doc (no distillation)
0.323
—
—
SPLADE-cocondenser
0.340
125
~0.11 MB
Dense 768d
—
768
3.07 MB
v12b matches SPLADE's sparsity (139 vs 125 NNZ) at 86% of v14's quality. Index is 25.6x smaller than dense vectors. Note: v12b is 10% behind Pyserini BM25 (0.325).
Special Property
v12b holds the best-ever NDCG@10 at threshold=0.0 (no sparsity filtering): 0.298. The model's raw retrieval quality is excellent — the low NNZ at t=0.3 means aggressive thresholding removes some useful terms.
v12b is also the foundation checkpoint for v14. It represents the "backbone quality" that v14 built on.
When to Use v12b Over v14
Index size matters (2.6x smaller than v14)
Query latency matters (fewer posting list lookups)
You need SPLADE-level sparsity without distillation
You're building a system with quality/efficiency tradeoff
Training
Foundation: fles1-v7 (1 generation from bert-base-uncased)
Data: 200,000 MS MARCO random negatives
Epochs: 2 (12,500 steps)
Loss: InfoNCE (τ=0.05) + L2 FLOPS (λ_d=0.00003) + anti-collapse
Controller: Step-interval CLFR, adjusted every ~6,250 steps (target_nnz_d=400, gain=0.1)
Epoch 1 → Epoch 2: λ_d adjusted from 0.000030 to 0.000034 (+13%)
Optimizer: AdamW, lr=2e-5, batch_size=32, 7 negatives
Hardware: 1× A100 80GB, ~2 hours
Note: v12b was trained with L2 FLOPS (before the L1 discovery). The L2 training produced more aggressive sparsity (NNZ=139) which is an advantage for efficiency.
The Step-Interval CLFR Proof
v12b is the proof that step-interval control works:
Approach
Adjustments
NDCG@10
No control (v12a)
0
0.276
Per-step (v11c)
12,500
0.297
Epoch-level (v12b)
1
0.292
One adjustment. +5.8% over no control. Within 1.7% of 12,500 chaotic adjustments.
Golvis Tavarez — Mindoval, Inc.We thank Microsoft Corporation for supporting this research through the Microsoft for Startups program.https://mindoval.com/ai-research
