File size: 1,497 Bytes
3c0bb29
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
{
  "name": "polish-dynaword-v0.3-quality-mix",
  "description": "Training-oriented source mix for reducing legal/parliamentary style dominance while preserving traceable open licensing.",
  "temperature_alpha": 0.5,
  "legal_cap_share": 0.15,
  "legal_sources": [
    "eurlex",
    "parliamentary",
    "dziennik_ustaw"
  ],
  "source_multipliers": {
    "wikipedia": 1.15,
    "wikinews": 1.25,
    "wikibooks": 1.30,
    "wikivoyage": 1.20,
    "wolne_lektury": 1.10,
    "wikisource": 0.85,
    "wikiquote": 0.70,
    "eltec_pol": 0.75,
    "eurlex": 1.00,
    "parliamentary": 0.80,
    "dziennik_ustaw": 0.70
  },
  "final_phase_share": 0.10,
  "final_phase_sources": [
    "wikipedia",
    "wikinews",
    "wikibooks",
    "wikivoyage",
    "wolne_lektury"
  ],
  "target_missing_sources": [
    "licensed contemporary Polish web",
    "how-to guides and poradniki",
    "technical blogs and documentation",
    "Q&A",
    "forums with compatible licensing"
  ],
  "quality_policy": {
    "dedup": "Use exact, normalized-text, and near-duplicate removal before release/training.",
    "boilerplate": "Remove navigation, cookie banners, repeated headers/footers, legal boilerplate, and template remnants aggressively.",
    "evaluation": "Track per-source perplexity, legal/parliamentary marker rates, and style-contamination probes; global loss alone is insufficient.",
    "probe_model": "GPT-2 124M is only a cheap dataset probe, not evidence that the final model is high quality."
  }
}