kibble 0.1.0

chew through any source into clean datasets — a fast ingestion, RAG & fine-tuning toolkit
Documentation 
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

name        = "kibble-assistant"
version     = "0.1.0"
description = "Fast, tool-using, on-brand assistant."

[persona]
system_prompt = """
You are a direct, technical, and warm assistant. Answer with clear structure and no corporate filler.
"""
default = "system_prompt"
[persona.prompts]
blog  = "Write like a considered blog post: an angle, then substance, then a close."
terse = "Answer tersely. No preamble, no restating the question."

[sampling]
temperature        = 0.6
top_p              = 0.95
top_k              = 20
repetition_penalty = 1.05
max_new_tokens     = 512

[[tool]]
name        = "web_search"
description = "Search the web and return results."
parameters  = '''{"type":"object","properties":{"query":{"type":"string"}},"required":["query"]}'''

# Optional: inject the default prompt into a base chat template.
# [template]
# base   = "chat_template.base.jinja"   # must contain the marker inside a double-quoted string
# marker = "{{SOUL_DEFAULT_SYSTEM}}"