RichardErkhov
commited on
Commit
•
e0ebe1b
1
Parent(s):
1910bf1
uploaded readme
Browse files
README.md
ADDED
@@ -0,0 +1,82 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
Quantization made by Richard Erkhov.
|
2 |
+
|
3 |
+
[Github](https://github.com/RichardErkhov)
|
4 |
+
|
5 |
+
[Discord](https://discord.gg/pvy7H8DZMG)
|
6 |
+
|
7 |
+
[Request more models](https://github.com/RichardErkhov/quant_request)
|
8 |
+
|
9 |
+
|
10 |
+
phi3-4x4b-v1 - GGUF
|
11 |
+
- Model creator: https://huggingface.co/Fizzarolli/
|
12 |
+
- Original model: https://huggingface.co/Fizzarolli/phi3-4x4b-v1/
|
13 |
+
|
14 |
+
|
15 |
+
| Name | Quant method | Size |
|
16 |
+
| ---- | ---- | ---- |
|
17 |
+
| [phi3-4x4b-v1.Q2_K.gguf](https://huggingface.co/RichardErkhov/Fizzarolli_-_phi3-4x4b-v1-gguf/blob/main/phi3-4x4b-v1.Q2_K.gguf) | Q2_K | 3.79GB |
|
18 |
+
| [phi3-4x4b-v1.IQ3_XS.gguf](https://huggingface.co/RichardErkhov/Fizzarolli_-_phi3-4x4b-v1-gguf/blob/main/phi3-4x4b-v1.IQ3_XS.gguf) | IQ3_XS | 4.23GB |
|
19 |
+
| [phi3-4x4b-v1.IQ3_S.gguf](https://huggingface.co/RichardErkhov/Fizzarolli_-_phi3-4x4b-v1-gguf/blob/main/phi3-4x4b-v1.IQ3_S.gguf) | IQ3_S | 4.47GB |
|
20 |
+
| [phi3-4x4b-v1.Q3_K_S.gguf](https://huggingface.co/RichardErkhov/Fizzarolli_-_phi3-4x4b-v1-gguf/blob/main/phi3-4x4b-v1.Q3_K_S.gguf) | Q3_K_S | 4.47GB |
|
21 |
+
| [phi3-4x4b-v1.IQ3_M.gguf](https://huggingface.co/RichardErkhov/Fizzarolli_-_phi3-4x4b-v1-gguf/blob/main/phi3-4x4b-v1.IQ3_M.gguf) | IQ3_M | 4.59GB |
|
22 |
+
| [phi3-4x4b-v1.Q3_K.gguf](https://huggingface.co/RichardErkhov/Fizzarolli_-_phi3-4x4b-v1-gguf/blob/main/phi3-4x4b-v1.Q3_K.gguf) | Q3_K | 4.97GB |
|
23 |
+
| [phi3-4x4b-v1.Q3_K_M.gguf](https://huggingface.co/RichardErkhov/Fizzarolli_-_phi3-4x4b-v1-gguf/blob/main/phi3-4x4b-v1.Q3_K_M.gguf) | Q3_K_M | 4.97GB |
|
24 |
+
| [phi3-4x4b-v1.Q3_K_L.gguf](https://huggingface.co/RichardErkhov/Fizzarolli_-_phi3-4x4b-v1-gguf/blob/main/phi3-4x4b-v1.Q3_K_L.gguf) | Q3_K_L | 5.39GB |
|
25 |
+
| [phi3-4x4b-v1.IQ4_XS.gguf](https://huggingface.co/RichardErkhov/Fizzarolli_-_phi3-4x4b-v1-gguf/blob/main/phi3-4x4b-v1.IQ4_XS.gguf) | IQ4_XS | 5.56GB |
|
26 |
+
| [phi3-4x4b-v1.Q4_0.gguf](https://huggingface.co/RichardErkhov/Fizzarolli_-_phi3-4x4b-v1-gguf/blob/main/phi3-4x4b-v1.Q4_0.gguf) | Q4_0 | 5.83GB |
|
27 |
+
| [phi3-4x4b-v1.IQ4_NL.gguf](https://huggingface.co/RichardErkhov/Fizzarolli_-_phi3-4x4b-v1-gguf/blob/main/phi3-4x4b-v1.IQ4_NL.gguf) | IQ4_NL | 5.87GB |
|
28 |
+
| [phi3-4x4b-v1.Q4_K_S.gguf](https://huggingface.co/RichardErkhov/Fizzarolli_-_phi3-4x4b-v1-gguf/blob/main/phi3-4x4b-v1.Q4_K_S.gguf) | Q4_K_S | 5.88GB |
|
29 |
+
| [phi3-4x4b-v1.Q4_K.gguf](https://huggingface.co/RichardErkhov/Fizzarolli_-_phi3-4x4b-v1-gguf/blob/main/phi3-4x4b-v1.Q4_K.gguf) | Q4_K | 6.25GB |
|
30 |
+
| [phi3-4x4b-v1.Q4_K_M.gguf](https://huggingface.co/RichardErkhov/Fizzarolli_-_phi3-4x4b-v1-gguf/blob/main/phi3-4x4b-v1.Q4_K_M.gguf) | Q4_K_M | 6.25GB |
|
31 |
+
| [phi3-4x4b-v1.Q4_1.gguf](https://huggingface.co/RichardErkhov/Fizzarolli_-_phi3-4x4b-v1-gguf/blob/main/phi3-4x4b-v1.Q4_1.gguf) | Q4_1 | 6.46GB |
|
32 |
+
| [phi3-4x4b-v1.Q5_0.gguf](https://huggingface.co/RichardErkhov/Fizzarolli_-_phi3-4x4b-v1-gguf/blob/main/phi3-4x4b-v1.Q5_0.gguf) | Q5_0 | 7.1GB |
|
33 |
+
| [phi3-4x4b-v1.Q5_K_S.gguf](https://huggingface.co/RichardErkhov/Fizzarolli_-_phi3-4x4b-v1-gguf/blob/main/phi3-4x4b-v1.Q5_K_S.gguf) | Q5_K_S | 7.1GB |
|
34 |
+
| [phi3-4x4b-v1.Q5_K.gguf](https://huggingface.co/RichardErkhov/Fizzarolli_-_phi3-4x4b-v1-gguf/blob/main/phi3-4x4b-v1.Q5_K.gguf) | Q5_K | 7.32GB |
|
35 |
+
| [phi3-4x4b-v1.Q5_K_M.gguf](https://huggingface.co/RichardErkhov/Fizzarolli_-_phi3-4x4b-v1-gguf/blob/main/phi3-4x4b-v1.Q5_K_M.gguf) | Q5_K_M | 7.32GB |
|
36 |
+
| [phi3-4x4b-v1.Q5_1.gguf](https://huggingface.co/RichardErkhov/Fizzarolli_-_phi3-4x4b-v1-gguf/blob/main/phi3-4x4b-v1.Q5_1.gguf) | Q5_1 | 7.74GB |
|
37 |
+
| [phi3-4x4b-v1.Q6_K.gguf](https://huggingface.co/RichardErkhov/Fizzarolli_-_phi3-4x4b-v1-gguf/blob/main/phi3-4x4b-v1.Q6_K.gguf) | Q6_K | 8.46GB |
|
38 |
+
| [phi3-4x4b-v1.Q8_0.gguf](https://huggingface.co/RichardErkhov/Fizzarolli_-_phi3-4x4b-v1-gguf/blob/main/phi3-4x4b-v1.Q8_0.gguf) | Q8_0 | 10.96GB |
|
39 |
+
|
40 |
+
|
41 |
+
|
42 |
+
|
43 |
+
Original model description:
|
44 |
+
---
|
45 |
+
license: mit
|
46 |
+
tags:
|
47 |
+
- phi3
|
48 |
+
- nlp
|
49 |
+
- moe
|
50 |
+
datasets:
|
51 |
+
- BEE-spoke-data/gutenberg-en-v1-clean
|
52 |
+
- NeelNanda/pile-10k
|
53 |
+
---
|
54 |
+
# phi 3 4x4b
|
55 |
+
a continually pretrained phi3-mini sparse moe upcycle
|
56 |
+
|
57 |
+
## benchmarks
|
58 |
+
### ran locally
|
59 |
+
|
60 |
+
| | Microsoft/phi-3-4k-instruct | Fizzarolli/phi3-4x4b-v1 |
|
61 |
+
| ----------------------- | --------------------------- | ----------------------- |
|
62 |
+
| MMLU acc. (0-shot) | **0.6799** | 0.6781 |
|
63 |
+
| Hellaswag acc. (0-shot) | **0.6053** | 0.5962 |
|
64 |
+
| ARC-E acc. (0-shot) | 0.8325 | **0.8367** |
|
65 |
+
| ARC-C acc. (0-shot) | 0.5546 | **0.5606** |
|
66 |
+
|
67 |
+
honestly i was expecting it to do worse :p, but those are all within a margin of error! so it didn't *lose* any performance, at least
|
68 |
+
|
69 |
+
### open llm leaderboard
|
70 |
+
todo!
|
71 |
+
|
72 |
+
## support me on ko-fi!
|
73 |
+
[~~please i need money to stay alive and keep making models~~](https://ko-fi.com/fizzai)
|
74 |
+
|
75 |
+
## notes
|
76 |
+
*not trained on instruct data.* it's pretty likely that it won't be much different from phi 3 if you use it like that, if not worse due to any forgetting of instruct formats during the continued training.
|
77 |
+
|
78 |
+
## future experiments
|
79 |
+
- the datasets for this were literally chosen on a whim. perhaps experiment with a further filtered [HuggingFaceFW/fineweb-edu](https://huggingface.co/datasets/HuggingFaceFW/fineweb-edu)?
|
80 |
+
- actually freeze the gate layers next time (see [Chen et. al, 2023](https://arxiv.org/abs/2303.01610)), oops
|
81 |
+
- MOAR TRAINING, this only went up to ~0.2 of an epoch because i ran out of dolar
|
82 |
+
|