Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,51 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: mit
|
| 3 |
+
tags:
|
| 4 |
+
- sparseflow
|
| 5 |
+
- sparse-attention
|
| 6 |
+
- efficient-nlp
|
| 7 |
+
datasets:
|
| 8 |
+
- gsm8k
|
| 9 |
+
- lighteval/MATH
|
| 10 |
+
- allenai/ai2_arc
|
| 11 |
+
- tau/commonsense_qa
|
| 12 |
+
- piqa
|
| 13 |
+
- allenai/sciq
|
| 14 |
+
- trivia_qa
|
| 15 |
+
- nq_open
|
| 16 |
+
- wikitext
|
| 17 |
+
---
|
| 18 |
+
|
| 19 |
+
# SparseFlow v8
|
| 20 |
+
|
| 21 |
+
Efficient language model with **sparse attention** and **persistent memory**.
|
| 22 |
+
|
| 23 |
+
## 📊 REAL Measured Metrics
|
| 24 |
+
|
| 25 |
+
| Metric | Value |
|
| 26 |
+
|--------|-------|
|
| 27 |
+
| Parameters | 71,359,746 |
|
| 28 |
+
| Perplexity | 14.77 |
|
| 29 |
+
| Attention Sparsity | 87.5% |
|
| 30 |
+
| Channel Sparsity | 75.0% |
|
| 31 |
+
| Peak Memory | 3.67 GB |
|
| 32 |
+
|
| 33 |
+
## 🏗️ Architecture
|
| 34 |
+
|
| 35 |
+
- **Sparse Token Attention**: Attends to top-64 tokens per position
|
| 36 |
+
- **Sparse Channel FFN**: Activates top-128 channels
|
| 37 |
+
- **Persistent Memory**: 20,000 memory vectors
|
| 38 |
+
- **8 Transformer layers** with 512 dim
|
| 39 |
+
|
| 40 |
+
## 📚 Training Data
|
| 41 |
+
|
| 42 |
+
Open source datasets only:
|
| 43 |
+
- GSM8K, MATH (mathematics)
|
| 44 |
+
- ARC, OpenBookQA, SciQ (science & reasoning)
|
| 45 |
+
- CommonsenseQA, PIQA (common sense)
|
| 46 |
+
- TriviaQA, Natural Questions (factual)
|
| 47 |
+
- WikiText-103 (language modeling)
|
| 48 |
+
|
| 49 |
+
## 👨💻 Author
|
| 50 |
+
|
| 51 |
+
**Logo (Mike Amega)** — Ame Web Studio
|