Document exact generation provenance

#2
Files changed (1) hide show
  1. README.md +29 -1
README.md CHANGED
@@ -18,7 +18,7 @@ Binary corpus for detecting **present-day LLM-generated fake product reviews**.
18
 
19
  - **OR (genuine)**: real human Amazon-style reviews (the authentic half of the
20
  Salminen et al. Kaggle "Fake Reviews" dataset).
21
- - **CG (fake)**: fresh fake reviews generated by **deepseek-v4-pro**. Each fake is
22
  *grounded on a paired real review* (same category and star rating, length
23
  matched) and prompted to be a different, believable review for the same
24
  product — so category/rating/length marginals match the genuine half and the
@@ -35,6 +35,34 @@ Both halves are normalised to a single ASCII punctuation convention (smart
35
  quotes/em-dashes -> ASCII) so Unicode is not a spurious tell. Exact and
36
  cross-class duplicates removed; classes balanced 1:1.
37
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
38
  ## Splits
39
  `train` / `validation` / `test` reproduce the frozen protocol split (seed 1998:
40
  80/20, then the 20% is split 30/70 into validation/test).
 
18
 
19
  - **OR (genuine)**: real human Amazon-style reviews (the authentic half of the
20
  Salminen et al. Kaggle "Fake Reviews" dataset).
21
+ - **CG (fake)**: fresh fake reviews generated by **DeepSeek-v4-pro**. Each fake is
22
  *grounded on a paired real review* (same category and star rating, length
23
  matched) and prompted to be a different, believable review for the same
24
  product — so category/rating/length marginals match the genuine half and the
 
35
  quotes/em-dashes -> ASCII) so Unicode is not a spurious tell. Exact and
36
  cross-class duplicates removed; classes balanced 1:1.
37
 
38
+ ## Generation protocol and provenance
39
+ Each generated review is grounded on the first 400 characters of a genuine
40
+ review. Source order uses seed 1998, and the review angle uses an RNG seeded by
41
+ the source-row index. Generation allows up to five attempts with temperatures
42
+ 0.85, 0.91, 0.97, 1.03, and 1.09; `top_p=0.95`. Accepted outputs contain at
43
+ least three words, stay within the recorded proportional length ceiling, and
44
+ have source/generated word-set Jaccard similarity below 0.6.
45
+
46
+ No API decoding seed was supplied. Provider model aliases may be mutable, and
47
+ some provider-side revisions were not recorded. The committed raw CSVs are the
48
+ canonical artifacts. Exact system/user prompts, commands, artifact SHA-256
49
+ values, generation history, and known unknowns are recorded in
50
+ [`DATA_PROVENANCE.md`](https://github.com/PurplexyFlower/Fake-review/blob/main/DATA_PROVENANCE.md)
51
+ and
52
+ [`dataset/provenance.json`](https://github.com/PurplexyFlower/Fake-review/blob/main/dataset/provenance.json).
53
+
54
+ ### This artifact
55
+
56
+ - Request model: `deepseek-v4-pro`
57
+ - Provider: DeepSeek API
58
+ - Provider endpoint: `https://api.deepseek.com`
59
+ - Immutable provider revision: not retained
60
+ - Raw artifact: `dataset/sota_fakes_deepseek.csv` (20,216 rows)
61
+ - Raw SHA-256: `a47568d219a38f5cc16207269d9d855ae370405fd703ae5c7d27e2131a019cc1`
62
+ - Final artifact: `dataset/modern_reviews_deepseek.csv` (40,424 rows; OR=20,212, CG=20,212)
63
+ - Final SHA-256: `c21b3a596101d488954058f56d72f35512a34dac1526ef8cbbd3d307fb005783`
64
+ - Artifact commit: `a1a77c419a59658f6cf06ea5c27e4f0bdab25229`
65
+
66
  ## Splits
67
  `train` / `validation` / `test` reproduce the frozen protocol split (seed 1998:
68
  80/20, then the 20% is split 30/70 into validation/test).