File size: 8,883 Bytes
45444a8 801cf6d 5555ddd 801cf6d 5555ddd 45444a8 2385765 b7577f8 2385765 b7577f8 250c782 801cf6d 812c543 2385765 812c543 2385765 812c543 2385765 812c543 2385765 812c543 801cf6d b7577f8 4aebaf7 d385b66 38e1149 b7577f8 5d748fc ac4bde9 5d748fc 5555ddd |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 |
---
language:
- de
license: apache-2.0
tags:
- hermeo
- laser
datasets:
- LeoLM/OpenSchnabeltier
pipeline_tag: conversational
model-index:
- name: germeo-7b-laser
results:
- task:
type: text-generation
name: Text Generation
dataset:
name: AI2 Reasoning Challenge (25-Shot)
type: ai2_arc
config: ARC-Challenge
split: test
args:
num_few_shot: 25
metrics:
- type: acc_norm
value: 60.75
name: normalized accuracy
source:
url: https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=aari1995/germeo-7b-laser
name: Open LLM Leaderboard
- task:
type: text-generation
name: Text Generation
dataset:
name: HellaSwag (10-Shot)
type: hellaswag
split: validation
args:
num_few_shot: 10
metrics:
- type: acc_norm
value: 82.81
name: normalized accuracy
source:
url: https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=aari1995/germeo-7b-laser
name: Open LLM Leaderboard
- task:
type: text-generation
name: Text Generation
dataset:
name: MMLU (5-Shot)
type: cais/mmlu
config: all
split: test
args:
num_few_shot: 5
metrics:
- type: acc
value: 60.57
name: accuracy
source:
url: https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=aari1995/germeo-7b-laser
name: Open LLM Leaderboard
- task:
type: text-generation
name: Text Generation
dataset:
name: TruthfulQA (0-shot)
type: truthful_qa
config: multiple_choice
split: validation
args:
num_few_shot: 0
metrics:
- type: mc2
value: 53.83
source:
url: https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=aari1995/germeo-7b-laser
name: Open LLM Leaderboard
- task:
type: text-generation
name: Text Generation
dataset:
name: Winogrande (5-shot)
type: winogrande
config: winogrande_xl
split: validation
args:
num_few_shot: 5
metrics:
- type: acc
value: 75.61
name: accuracy
source:
url: https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=aari1995/germeo-7b-laser
name: Open LLM Leaderboard
- task:
type: text-generation
name: Text Generation
dataset:
name: GSM8k (5-shot)
type: gsm8k
config: main
split: test
args:
num_few_shot: 5
metrics:
- type: acc
value: 43.37
name: accuracy
source:
url: https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=aari1995/germeo-7b-laser
name: Open LLM Leaderboard
---
(Evaluation WIP)
## Hermes + Leo + German Laser = Germeo
## Germeo-7B-Laser
A German-English understanding, but German-only speaking model merged from Hermeo-7B.
### Model details
**Merged from**: leo-mistral-hessianai-7b-chat and DPOpenHermes-7B-v2
**Model type**: Causal decoder-only transformer language model
**Languages**: German replies with English Understanding Capabilities
**Laser-Data**: LeoLM/OpenSchnabeltier
This is an early experiment on laser and its influence on language understanding. It generally improves the language understanding capabilities.
The hypothesis is that it degrades the probability of English replies and increasing those of German replies. The models internal German capabilities are boosted.
Will keep you updated..
### Acknowledgements:
I would like to thank everyone that participated in making this model and its training possible:
To [@malteos](https://huggingface.co/malteos) for hermeo
To [@cognitivecomputations](https://huggingface.co/cognitivecomputations) and Fernando Fernandes Neto for their implementation of LASER
To [@LeoLM](https://huggingface.co/LeoLM) and Björn for the OpenSchnabeltier dataset.
### Prompt format:
```python
streamer = TextStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)
# Convert prompt to tokens
prompt_template = """<|im_start|>system
Du bist ein hilfreicher Assistent.<|im_end|>
<|im_start|>user
{prompt}<|im_end|>
<|im_start|>assistant"""
prompt = "Schreibe eine Stellenanzeige für Data Scientist bei AXA!"
final_prompt = prompt_template.format(prompt=prompt)
```
#### Limit the model to output reply-only:
To solve this, you need to implement a custom stopping criteria:
```python
from transformers import StoppingCriteria
class GermeoStoppingCriteria(StoppingCriteria):
def __init__(self, target_sequence, prompt):
self.target_sequence = target_sequence
self.prompt=prompt
def __call__(self, input_ids, scores, **kwargs):
# Get the generated text as a string
generated_text = tokenizer.decode(input_ids[0])
generated_text = generated_text.replace(self.prompt,'')
# Check if the target sequence appears in the generated text
if self.target_sequence in generated_text:
return True # Stop generation
return False # Continue generation
def __len__(self):
return 1
def __iter__(self):
yield self
```
This then expects your input prompt (formatted as given into the model), and a stopping criteria, in this case the im_end token. Simply add it to the generation:
```python
generation_output = model.generate(
tokens,
streamer=streamer,
max_new_tokens=1012,
stopping_criteria=GermeoStoppingCriteria("<|im_end|>", prompt_template.format(prompt=prompt))
)
```
### German benchmarks
| **German tasks:** | **MMLU-DE** | **Hellaswag-DE** | **ARC-DE** |**Average** |
|-------------------------------|-------------|---------------|--------------|--------------|
| **Models / Few-shots:** | _(5 shots)_ | _(10 shots)_ | _(24 shots)_ | |
| _7B parameters_ | | | | |
| llama-2-7b | 0.400 | 0.513 | 0.381 | 0.431 |
| leo-hessianai-7b | 0.400 | 0.609 | 0.429 | 0.479 |
| bloom-6b4-clp-german | 0.274 | 0.550 | 0.351 | 0.392 |
| mistral-7b | **0.524** | 0.588 | 0.473 | 0.528 |
| leo-mistral-hessianai-7b | 0.481 | 0.663 | 0.485 | 0.543 |
| leo-mistral-hessianai-7b-chat | 0.458 | 0.617 | 0.465 | 0.513 |
| DPOpenHermes-7B-v2 | 0.517 | 0.603 | 0.515 | 0.545 |
| hermeo-7b | 0.511 | **0.668** | **0.528** | **0.569** |
| **germeo-7b-laser (this model)**| ? | ? | ? | ? |
| _13B parameters_ | | | | |
| llama-2-13b | 0.469 | 0.581 | 0.468 | 0.506 |
| leo-hessianai-13b | **0.486** | **0.658** | **0.509** | **0.551** |
| _70B parameters_ | | | | |
| llama-2-70b | 0.597 | 0.674 | 0.561 | 0.611 |
| leo-hessianai-70b | **0.653** | **0.721** | **0.600** | **0.658** |
Even though the model does not generate English text without being explicitly asked, performance on English Benchmarks is still up:
### English benchmarks
| **English tasks:** | **MMLU** | **Hellaswag** | **ARC** | **Average** |
|------------------------------------|-------------|---------------|--------------|-------------|
| **Models / Few-shots:** | _(5 shots)_ | _(10 shots)_ | _(24 shots)_ | |
| llama-2-7b | 0.466 | 0.786 | 0.530 | 0.594 |
| leolm-hessianai-7b | 0.423 | 0.759 | 0.522 | 0.568 |
| bloom-6b4-clp-german | 0.264 | 0.525 | 0.328 | 0.372 |
| mistral-7b | **0.635** | **0.832** | 0.607 | **0.691** |
| leolm-mistral-hessianai-7b | 0.550 | 0.777 | 0.518 | 0.615 |
| hermeo-7b | 0.601 | 0.821 | **0.620** | 0.681 |
| germeo-7b-laser (this model) | 0.601 | 0.828 | 0.608 | 0.679 |
# [Open LLM Leaderboard Evaluation Results](https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard)
Detailed results can be found [here](https://huggingface.co/datasets/open-llm-leaderboard/details_aari1995__germeo-7b-laser)
| Metric |Value|
|---------------------------------|----:|
|Avg. |62.82|
|AI2 Reasoning Challenge (25-Shot)|60.75|
|HellaSwag (10-Shot) |82.81|
|MMLU (5-Shot) |60.57|
|TruthfulQA (0-shot) |53.83|
|Winogrande (5-shot) |75.61|
|GSM8k (5-shot) |43.37|
|