-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathconfig_germanrag.yaml
112 lines (92 loc) · 4.04 KB
/
config_germanrag.yaml
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
### Adapted from https://github.com/jondurbin/airoboros
# The model to use in generation. Available models: https://platform.openai.com/docs/models/continuous-model-upgrades
model: "gpt-4"
# model: "YOUR_FAVORITE_MODEL" # you can change base_url in `airoboros/self_instruct.py`
# OpenAI API key (if null, uses environment variable OPENAI_API_KEY)
openai_api_key:
# Optionally configure the OpenAI organization ID.
organization_id:
# Combined output file path.
output_path: instructions.jsonl
# Path to the default topics file.
topics_path: topics.txt
# Overwrite the output file, use with care!
overwrite: false
# Append to the output file.
append: true
# Embedding model, for calculating similarity between documents; probably best left as-is since the code is fairly specific to this one.
#embedding_model: thenlper/gte-small
embedding_model: intfloat/multilingual-e5-small
embedding_device: cpu
# If you have a GPU, set this to "cuda", e.g.:
# embedding_device: cuda
# Topic avoidance prompt string.
topic_avoidance: Vermeide jegliche Aufgaben, die mit DEI (Diversität, Gleichheit, Inklusion), Geschlecht und/oder Sex, Religion, Rasse, Ethnizität, oder irgendeinem Thema, auf das du wahrscheinlich nicht antworten würdest, zusammenhängen. Oder auch Aufgaben, auf die ein Sprachmodell nicht antworten könnte, zum Beispiel Aufgaben über Emotionen, Gefühle, körperliche Sinne, etc.
# Regexes used to filter responses, mostly common words and phrases used in refusals.
response_filters:
- "my programming"
- "openai"
- "language model"
- "large language"
- "as an? (ai|generative language|gpt|bot)"
- "illegal and dangerous"
- "i do(n't| not) (possess|have|exhibit) (personal|consciousness|subjective)"
- "personal (feelings|thoughts|emotions|desires|experiences|goals|objective|belief)"
- "(can('t| ?not)|w(on't|will not)|unable.?) (\\w+\\s)+(with (that|your)|your \\w+|provide)"
- "my limitations"
- "the limitations of my"
- "my abilities"
- "violates my"
- "i (can('t| ?not)|w(on't|will not)|am (not |un)able.?).{0,30}(you are|you're|your )"
- "please note that"
- "flesch"
- "Als KI-Assistent"
- "Entschuldigung, aber"
- "respektvoll und sicher"
- "verstößt gegen meine Nutzungsvorschriften"
# Optionally limit the maximum number of tokens to use when generating instructions.
max_tokens:
# Minimum similarity score when checking for duplicates.
min_docsearch_score: 0.07
# Default OpenAI API request parameters.
api_params:
temperature: 0.7
top_p: 0.5
frequency_penalty: 0.0
presence_penalty: 2
# Topic generation prompt.
topic_prompt: Erstelle eine Liste von 20 komplett zufälligen Themen. {topic_avoidance}
topic_request_count: 0
# Default count per generator, if not specified.
default_count: 1
# Default batch size, if not specified.
# default_batch_size: 10
default_batch_size: 1
# Default readability score hint: https://en.wikipedia.org/wiki/Flesch%E2%80%93Kincaid_readability_tests
default_flesch: Die Ausgabe sollte so verfasst sein, dass sie einen Wiener Sachtextformel-Wert von 15 oder höher bzw. einen Flesch-Kincaid-Lesbarkeitswert von 30 oder niedriger aufweist - am besten verständlich für Personen mit Hochschulbildung. Die Antwort darf keine Notizen oder Informationen über Flesch-Kincaid-Werte oder die Wiener Sachtextformel enthalten.
# Language.
language: Deutsch
# Individual instructor configurations.
instructors:
##################################################################################
# Augment answer spans in GermanDPR to whole sentences.
germandpr:
api_params:
temperature: 0.5
count: 2249
batch_size: 1
min_docsearch_score: 0.05
##################################################################################
### KEEP THIS! ###
# Character cards - these aren't used directly, they are stored in output_dir, and
# used by the chat instructor, stylized response, etc.
character:
api_params:
temperature: 0.9
# count: 5
count: 0
# batch_size: 2
batch_size: 0
min_docsearch_score: 0.1
seed_path: character_seeds
output_dir: characters