Initial BiBoP release (private)
Browse files- README.md +15 -0
- config.json +14 -0
- model.pt +3 -0
- tokenizer.json +0 -0
README.md
ADDED
|
@@ -0,0 +1,15 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: mit
|
| 3 |
+
tags: [ternary, bitnet, bibop, language-model]
|
| 4 |
+
---
|
| 5 |
+
# BiBoP-91M
|
| 6 |
+
|
| 7 |
+
Latent-free **ternary** language model trained with **BiBoP** (weights in {-1,0,+1}, no latent
|
| 8 |
+
full-precision copy; only a bf16 flip momentum as optimiser state). Trained on a single RTX 3060 (6 GB).
|
| 9 |
+
|
| 10 |
+
- Data: TinyStories (~90M tokens)
|
| 11 |
+
- BLiMP macro-accuracy (own harness): **59.07**
|
| 12 |
+
- Arch: d768, 12 layers, 1 full-attention layer(s) + sliding-window, NoPE, tied interface.
|
| 13 |
+
- Recipe: per-row RMS flip, tau=(2.0,2.75); see the code/report at github.com/ValerioDolci/bibop.
|
| 14 |
+
|
| 15 |
+
Proof of concept; budget-scoped (not SOTA, no speed claim). Load with the `bibop` package (config.json + model.pt + tokenizer.json).
|
config.json
ADDED
|
@@ -0,0 +1,14 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"variant": "bop",
|
| 3 |
+
"d": 768,
|
| 4 |
+
"layers": 12,
|
| 5 |
+
"heads": 8,
|
| 6 |
+
"window": 64,
|
| 7 |
+
"n_full": 1,
|
| 8 |
+
"vocab": 8192,
|
| 9 |
+
"eot_id": 0,
|
| 10 |
+
"tau": [
|
| 11 |
+
2.0,
|
| 12 |
+
2.75
|
| 13 |
+
]
|
| 14 |
+
}
|
model.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:ec0981870b2a82545f0b7c0041efa9af5d08a7308a872af25e8ec95624b3b2e0
|
| 3 |
+
size 46657051
|
tokenizer.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|