Update README.md
Browse files
README.md
CHANGED
@@ -8,9 +8,20 @@ license: bigscience-openrail-m
|
|
8 |
---
|
9 |
|
10 |
|
11 |
-
GeoV-9B is a 20 billion parameter autoregressive language model
|
12 |
|
13 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
14 |
|
15 |
- Developed by: [Georges Harik](http://twitter.com/gharik)
|
16 |
- Model type: Transformer-based Language Model
|
@@ -29,3 +40,26 @@ GeoV-9B is a 20 billion parameter autoregressive language model
|
|
29 |
| Sequence Length | 2049 |
|
30 |
</figure>
|
31 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
8 |
---
|
9 |
|
10 |
|
11 |
+
[GeoV](https://huggingface.co/docs/transformers/model_doc/geov)-9B is a 20 billion parameter autoregressive language model.
|
12 |
|
13 |
+
The GeoV model was designed by Georges Harik and uses
|
14 |
+
[Rotary Positional Embeddings with Relative distances (RoPER)](http://research.labml.ai/RoPER.html)
|
15 |
+
by [Georges Hark](https://twitter.com/ghark) and [Varuna Jayasiri](https://twitter.com/vpj).
|
16 |
+
|
17 |
+
[RoPER]((http://research.labml.ai/RoPER.html),
|
18 |
+
in addition to using relative positions in the attention score calculation by RoPE embeddings,
|
19 |
+
adds relative positional information explicitly to value embeddings.
|
20 |
+
Specifically, it incorporates the relative positions of the tokens paid attention to.
|
21 |
+
RoPER gives better performance in algorithmic tasks.
|
22 |
+
Results have shown an improvement over RoPE in a language modeling setting on a 3 billion parameter transformer.
|
23 |
+
|
24 |
+
## Model details
|
25 |
|
26 |
- Developed by: [Georges Harik](http://twitter.com/gharik)
|
27 |
- Model type: Transformer-based Language Model
|
|
|
40 |
| Sequence Length | 2049 |
|
41 |
</figure>
|
42 |
|
43 |
+
|
44 |
+
## Generation
|
45 |
+
|
46 |
+
The `generate()` method can be used to generate text using GeoV model.
|
47 |
+
|
48 |
+
```python
|
49 |
+
>>> from transformers import GeoVForCausalLM, GeoVTokenizer
|
50 |
+
|
51 |
+
>>> model = GeoVForCausalLM.from_pretrained("GoeV/GeoV-9b")
|
52 |
+
>>> tokenizer = GeoVTokenizer.from_pretrained("GoeV/GeoV-9b")
|
53 |
+
|
54 |
+
>>> prompt = "In mathematics, topology is the study of"
|
55 |
+
|
56 |
+
>>> input_ids = tokenizer(prompt, return_tensors="pt").input_ids
|
57 |
+
|
58 |
+
>>> gen_tokens = model.generate(
|
59 |
+
... input_ids,
|
60 |
+
... do_sample=True,
|
61 |
+
... temperature=0.9,
|
62 |
+
... max_length=100,
|
63 |
+
... )
|
64 |
+
>>> gen_text = tokenizer.batch_decode(gen_tokens)[0]
|
65 |
+
```
|