Update README.md
Browse files
README.md
CHANGED
|
@@ -1,3 +1,70 @@
|
|
| 1 |
---
|
| 2 |
license: apache-2.0
|
| 3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
---
|
| 2 |
license: apache-2.0
|
| 3 |
---
|
| 4 |
+
|
| 5 |
+
Negative-v1.0 is a 67K-parameter small language model (SLM) featuring a custom architecture inspired by Needle2. Trained entirely on CPU over 600M tokens, Negative-v1.0 utilizes a byte-level tokenizer with 4 special tokens (<bos>, <eos>, <pad>, <unk>), resulting in a compact vocabulary size of 260.
|
| 6 |
+
|
| 7 |
+
## Architecture
|
| 8 |
+
|
| 9 |
+
Negative-v1.0 employs a compact, parameter-efficient architecture incorporating Engram memory, Hadamard FFNs with SwiGLU intervals, and an 8-stream topology powered by mHC.
|
| 10 |
+
|
| 11 |
+
- Vocab Size: `260`
|
| 12 |
+
- Max Position Embeddings: `96`
|
| 13 |
+
- Hidden Size: `32`
|
| 14 |
+
- Intermediate Size (for SwiGLU): `64`
|
| 15 |
+
- Total Number of Layers: `9`
|
| 16 |
+
- Hadamard Layers: `7`
|
| 17 |
+
- SwiGLU Layers: `2`
|
| 18 |
+
- Number of Heads: `4`
|
| 19 |
+
- Number of KV Heads: `2`
|
| 20 |
+
- Dimensions Per Head: `8`
|
| 21 |
+
- Use Per-Head Gating: `false`
|
| 22 |
+
- Use XSA: `false`
|
| 23 |
+
- Number of mHC Streams: `8`
|
| 24 |
+
- Use Engram: `true`
|
| 25 |
+
- Number of Engram Entries: `196`
|
| 26 |
+
- Engram Orders: `(4, 8)`
|
| 27 |
+
|
| 28 |
+
## Training Dataset
|
| 29 |
+
|
| 30 |
+
Negative was trained on 600 million tokens of a diverse dataset mixture comprising general web text, educational content, synthetic data, normalized code, and mathematics.
|
| 31 |
+
|
| 32 |
+
| Dataset | Share |
|
| 33 |
+
| :--- | :---: |
|
| 34 |
+
| **FineWeb-Edu** | 36.0% |
|
| 35 |
+
| **DCLM Baseline 1.0** | 22.9% |
|
| 36 |
+
| **FinePhrase** | 13.4% |
|
| 37 |
+
| **MGA FineWeb-Edu** | 10.3% |
|
| 38 |
+
| **Tiny Strange Textbooks** | 8.2% |
|
| 39 |
+
| **OpenMathInstruct-2** | 7.6% |
|
| 40 |
+
| **NPset-2 Python-Edu** | 1.6% |
|
| 41 |
+
|
| 42 |
+
## Benchmark Results
|
| 43 |
+
|
| 44 |
+
We benchmaked Negative-v1.0 on five tasks: Arc_Easy, Arc_Challenge, HellaSwag, PiQA, and ArithMark-3.0.
|
| 45 |
+
|
| 46 |
+
| Task | Metric | Score |
|
| 47 |
+
| :--- | :--- | :---: |
|
| 48 |
+
| **ARC Challenge** | `acc_norm` | 22.95% |
|
| 49 |
+
| **ARC Easy** | `acc_norm` | 27.65% |
|
| 50 |
+
| **HellaSwag** | `acc_norm` | 25.94% |
|
| 51 |
+
| **PIQA** | `acc_norm` | 49.62% |
|
| 52 |
+
| **ArithMark-3.0** | `acc_norm` | 31.50% |
|
| 53 |
+
|
| 54 |
+
Despite its compact size, Negative exhibits surprisingly competitive performance on knowledge-intensive and mathematical benchmarks within its parameter class.
|
| 55 |
+
|
| 56 |
+
## License
|
| 57 |
+
|
| 58 |
+
Apache 2.0.
|
| 59 |
+
|
| 60 |
+
## Citation
|
| 61 |
+
|
| 62 |
+
```
|
| 63 |
+
@misc{negative-v1.0,
|
| 64 |
+
title = {Negative-v1.0},
|
| 65 |
+
organization = {FromZero},
|
| 66 |
+
authors = {Paul Courneya},
|
| 67 |
+
year = {2026},
|
| 68 |
+
url = {https://huggingface.co/fromziro/Negative-v1.0]
|
| 69 |
+
}
|
| 70 |
+
```
|