Commit
•
68ee8b9
1
Parent(s):
f7d4ed0
Create README.md
Browse files
README.md
ADDED
@@ -0,0 +1,182 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
tags:
|
3 |
+
- merge
|
4 |
+
license: other
|
5 |
+
model-index:
|
6 |
+
- name: QuartetAnemoi-70B-t0.0001
|
7 |
+
results:
|
8 |
+
- task:
|
9 |
+
type: text-generation
|
10 |
+
name: Text Generation
|
11 |
+
dataset:
|
12 |
+
name: AI2 Reasoning Challenge (25-Shot)
|
13 |
+
type: ai2_arc
|
14 |
+
config: ARC-Challenge
|
15 |
+
split: test
|
16 |
+
args:
|
17 |
+
num_few_shot: 25
|
18 |
+
metrics:
|
19 |
+
- type: acc_norm
|
20 |
+
value: 73.38
|
21 |
+
name: normalized accuracy
|
22 |
+
source:
|
23 |
+
url: https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=alchemonaut/QuartetAnemoi-70B-t0.0001
|
24 |
+
name: Open LLM Leaderboard
|
25 |
+
- task:
|
26 |
+
type: text-generation
|
27 |
+
name: Text Generation
|
28 |
+
dataset:
|
29 |
+
name: HellaSwag (10-Shot)
|
30 |
+
type: hellaswag
|
31 |
+
split: validation
|
32 |
+
args:
|
33 |
+
num_few_shot: 10
|
34 |
+
metrics:
|
35 |
+
- type: acc_norm
|
36 |
+
value: 88.9
|
37 |
+
name: normalized accuracy
|
38 |
+
source:
|
39 |
+
url: https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=alchemonaut/QuartetAnemoi-70B-t0.0001
|
40 |
+
name: Open LLM Leaderboard
|
41 |
+
- task:
|
42 |
+
type: text-generation
|
43 |
+
name: Text Generation
|
44 |
+
dataset:
|
45 |
+
name: MMLU (5-Shot)
|
46 |
+
type: cais/mmlu
|
47 |
+
config: all
|
48 |
+
split: test
|
49 |
+
args:
|
50 |
+
num_few_shot: 5
|
51 |
+
metrics:
|
52 |
+
- type: acc
|
53 |
+
value: 75.42
|
54 |
+
name: accuracy
|
55 |
+
source:
|
56 |
+
url: https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=alchemonaut/QuartetAnemoi-70B-t0.0001
|
57 |
+
name: Open LLM Leaderboard
|
58 |
+
- task:
|
59 |
+
type: text-generation
|
60 |
+
name: Text Generation
|
61 |
+
dataset:
|
62 |
+
name: TruthfulQA (0-shot)
|
63 |
+
type: truthful_qa
|
64 |
+
config: multiple_choice
|
65 |
+
split: validation
|
66 |
+
args:
|
67 |
+
num_few_shot: 0
|
68 |
+
metrics:
|
69 |
+
- type: mc2
|
70 |
+
value: 69.53
|
71 |
+
source:
|
72 |
+
url: https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=alchemonaut/QuartetAnemoi-70B-t0.0001
|
73 |
+
name: Open LLM Leaderboard
|
74 |
+
- task:
|
75 |
+
type: text-generation
|
76 |
+
name: Text Generation
|
77 |
+
dataset:
|
78 |
+
name: Winogrande (5-shot)
|
79 |
+
type: winogrande
|
80 |
+
config: winogrande_xl
|
81 |
+
split: validation
|
82 |
+
args:
|
83 |
+
num_few_shot: 5
|
84 |
+
metrics:
|
85 |
+
- type: acc
|
86 |
+
value: 85.32
|
87 |
+
name: accuracy
|
88 |
+
source:
|
89 |
+
url: https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=alchemonaut/QuartetAnemoi-70B-t0.0001
|
90 |
+
name: Open LLM Leaderboard
|
91 |
+
- task:
|
92 |
+
type: text-generation
|
93 |
+
name: Text Generation
|
94 |
+
dataset:
|
95 |
+
name: GSM8k (5-shot)
|
96 |
+
type: gsm8k
|
97 |
+
config: main
|
98 |
+
split: test
|
99 |
+
args:
|
100 |
+
num_few_shot: 5
|
101 |
+
metrics:
|
102 |
+
- type: acc
|
103 |
+
value: 68.61
|
104 |
+
name: accuracy
|
105 |
+
source:
|
106 |
+
url: https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=alchemonaut/QuartetAnemoi-70B-t0.0001
|
107 |
+
name: Open LLM Leaderboard
|
108 |
+
---
|
109 |
+
|
110 |
+
<img src=https://huggingface.co/alchemonaut/QuartetAnemoi-70B-t0.0001/resolve/main/anemoi.png>
|
111 |
+
|
112 |
+
# QuartetAnemoi-70B-t0.0001
|
113 |
+
|
114 |
+
A sequential merge using a custom algorithm (NearSwap) of:
|
115 |
+
- [152334H/miqu-1-70b-sf](https://huggingface.co/152334H/miqu-1-70b-sf)
|
116 |
+
- [Sao10K/WinterGoddess-1.4x-70B-L2](https://huggingface.co/Sao10K/WinterGoddess-1.4x-70B-L2)
|
117 |
+
- [Aurora-Nights-70B-v1.0](https://huggingface.co/sophosympatheia/Aurora-Nights-70B-v1.0)
|
118 |
+
- [Xwin-LM-70B-V0.1](https://huggingface.co/Xwin-LM/Xwin-LM-70B-V0.1)
|
119 |
+
|
120 |
+
<br/>
|
121 |
+
|
122 |
+
In our testing, this model seems like a storyteller, as might be expected, but the changes from this merge are extremely soft. We were impressed that, unlike most models, at the end of a story it did not often use cliches such as "In the end", "And so", "beacon of hope", etc.
|
123 |
+
|
124 |
+
<br/>
|
125 |
+
|
126 |
+
This repo has the GGUF quants. Full weights are available at: [alchemonaut/QuartetAnemoi-70B-t0.0001](https://huggingface.co/alchemonaut/QuartetAnemoi-70B-t0.0001)
|
127 |
+
|
128 |
+
<br/>
|
129 |
+
<br/>
|
130 |
+
|
131 |
+
# NearSwap Algorithm
|
132 |
+
|
133 |
+
NearSwap retains most of the weights of the base model (Miqu), but when a weight is similar between the two, it is interpolated to the secondary model value. A parameter *t* specifies the sameness threshold. When the distance between two values is below *t*, the weight from the secondary model is used.
|
134 |
+
|
135 |
+
This version of the model uses *t* = 0.0001. At this *t*, about 0.8% of weights are fully switched to the secondary model during each pass. Model quality rapidly degrades above *t* = 0.0025:
|
136 |
+
|
137 |
+
- *t* = 0.0001 (~0.8% full swap): This model
|
138 |
+
- *t* = 0.0003 (~2% full swap)
|
139 |
+
- *t* = 0.001 (~10% full swap): [BoreanGale-70B](https://huggingface.co/alchemonaut/BoreanGale-70B)
|
140 |
+
- *t* = 0.0025 (~18% full swap): Generates one paragraph okay, but then reverts to garbage
|
141 |
+
- *t* = 0.005 (~35% full swap): Garbage; semi-related word lists
|
142 |
+
- *t* = 0.01 (~55% full swap): Garbage; pseudorandom tokens output
|
143 |
+
|
144 |
+
For QuartetAnemoi-70B-t0.0001, the three secondary models were each merged sequentially with *t* = 0.0001.
|
145 |
+
|
146 |
+
NearSwap implementation:
|
147 |
+
```
|
148 |
+
t: Union[float, np.ndarray],
|
149 |
+
v0: Union[np.ndarray, torch.Tensor],
|
150 |
+
v1: Union[np.ndarray, torch.Tensor],
|
151 |
+
...
|
152 |
+
lweight = numpy.absolute(v0-v1)
|
153 |
+
lweight = t / lweight
|
154 |
+
lweight = numpy.nan_to_num(lweight, nan=1.0, posinf=1.0, neginf=1.0)
|
155 |
+
numpy.clip(lweight, a_min=0.0, a_max=1.0, out=lweight)
|
156 |
+
res = lerp(lweight,v0,v1)
|
157 |
+
```
|
158 |
+
<br/>
|
159 |
+
<br/>
|
160 |
+
|
161 |
+
|
162 |
+
# License and Use
|
163 |
+
|
164 |
+
Since the ultimate origin of Miqu is at this time unknown beyond speculation, this model is for noncommercial research use only.
|
165 |
+
|
166 |
+
<br/>
|
167 |
+
<br/>
|
168 |
+
|
169 |
+
|
170 |
+
# [Open LLM Leaderboard Evaluation Results](https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard)
|
171 |
+
Detailed results can be found [here](https://huggingface.co/datasets/open-llm-leaderboard/details_alchemonaut__QuartetAnemoi-70B-t0.0001)
|
172 |
+
|
173 |
+
| Metric |Value|
|
174 |
+
|---------------------------------|----:|
|
175 |
+
|Avg. |76.86|
|
176 |
+
|AI2 Reasoning Challenge (25-Shot)|73.38|
|
177 |
+
|HellaSwag (10-Shot) |88.9|
|
178 |
+
|MMLU (5-Shot) |75.42|
|
179 |
+
|TruthfulQA (0-shot) |69.53|
|
180 |
+
|Winogrande (5-shot) |85.32|
|
181 |
+
|GSM8k (5-shot) |68.61|
|
182 |
+
|