LordofMonarchs commited on
Commit
4e052ee
·
verified ·
1 Parent(s): bb7583b

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +0 -79
README.md CHANGED
@@ -35,85 +35,6 @@ The pipeline is split into two phases. The offline phase has no time limit and p
35
  ![App Architecture](architecture.png)
36
 
37
 
38
- ```mermaid
39
- flowchart TD
40
- %% Define classes with light pastel fills, complementary borders, and dark text for contrast
41
- classDef data fill:#e0f2fe,stroke:#0284c7,stroke-width:1px,color:#0f172a;
42
- classDef llm fill:#f3e8ff,stroke:#7e22ce,stroke-width:1px,color:#0f172a;
43
- classDef offline fill:#ffedd5,stroke:#c2410c,stroke-width:1px,color:#0f172a;
44
- classDef online fill:#dcfce7,stroke:#15803d,stroke-width:1px,color:#0f172a;
45
-
46
- candidates_jsonl["candidates.jsonl<br/>(100,000 records)"]:::data
47
- submission_csv["submission.csv<br/>(100 ranked candidates)"]:::data
48
-
49
- subgraph offline_phase [OFFLINE PHASE]
50
- subgraph gemma3_annotation [GEMMA3 PAIRWISE ANNOTATION]
51
- stratified_sample["Stratified Sample of<br/>500 Candidates"]:::llm
52
- pairwise_comparisons["2,500 Pairwise Comparisons<br/>(Local Gemma3 LLM)"]:::llm
53
- elo_ratings["Laplace-smoothed<br/>Elo Ratings"]:::llm
54
- relevance_labels["Quartile Thresholding to<br/>Relevance Labels 0-3"]:::llm
55
-
56
- stratified_sample --> pairwise_comparisons
57
- pairwise_comparisons --> elo_ratings
58
- elo_ratings --> relevance_labels
59
- end
60
-
61
- bm25_index_build["BM25 Index Build<br/>(NumPy CSR matrix)"]:::offline
62
- static_feature_precompute["Static Feature Precompute<br/>(18 JD-independent features)"]:::offline
63
-
64
- lightgbm_training["LIGHTGBM TRAINING<br/>Objective: lambdarank<br/>early stop on NDCG@5"]:::offline
65
-
66
- bm25_index_build --> precomputed_artifacts_box
67
- static_feature_precompute --> precomputed_artifacts_box
68
- relevance_labels --> lightgbm_training
69
- static_feature_precompute --> lightgbm_training
70
- bm25_index_build --> lightgbm_training
71
-
72
- lightgbm_training --> precomputed_artifacts_box
73
-
74
- precomputed_artifacts_box["PRECOMPUTED ARTIFACTS<br/>lgbm_model.txt<br/>bm25_matrix.npz<br/>static_features.pkl"]:::data
75
- end
76
-
77
- subgraph online_ranking_phase [ONLINE RANKING PHASE]
78
- stage_0["Stage 0: Load Precomputed Artifacts<br/>(BM25, LightGBM, static features)"]:::online
79
-
80
- stage_1["Stage 1: Dual-pass BM25 Retrieval<br/>Pass A: JD skills on skills array<br/>Pass B: production keywords on career descriptions<br/>Narrow to ~8,500 candidates"]:::online
81
-
82
- stage_2["Stage 2: Load Stage 1 Records<br/>(Byte-offset index, O(1))"]:::online
83
-
84
- stage_2b["Stage 2b: Feature Engineering<br/>(22 features, adversarial detection,<br/>consistency score)"]:::online
85
-
86
- stage_4["Stage 4: LightGBM Inference<br/>(final_score = raw_score * consistency_score)"]:::online
87
-
88
- stage_5["Stage 5: Reasoning Compiler<br/>(Deterministic grammar, 4 templates,<br/>priority concerns, numeric audit)"]:::online
89
-
90
- stage_6["Stage 6: Blocking Audits + CSV Write<br/>(Honeypot, diversity, monotonicity checks)"]:::online
91
-
92
- stage_0 --> stage_1
93
- stage_1 --> stage_2
94
- stage_2 --> stage_2b
95
- stage_2b --> stage_4
96
- stage_4 --> stage_5
97
- stage_5 --> stage_6
98
- stage_6 --> submission_csv
99
- end
100
-
101
- candidates_jsonl --> offline_phase
102
- candidates_jsonl --> stage_1
103
-
104
- precomputed_artifacts_box --> stage_0
105
- precomputed_artifacts_box -.-> stage_2
106
- precomputed_artifacts_box -.-> stage_2b
107
- precomputed_artifacts_box -.-> stage_4
108
-
109
- %% Use transparent fills (none) so it inherits GitHub's native dark/light themes seamlessly
110
- style offline_phase fill:none,stroke:#777,stroke-width:2px,stroke-dasharray: 5 5
111
- style online_ranking_phase fill:none,stroke:#777,stroke-width:2px,stroke-dasharray: 5 5
112
- style gemma3_annotation fill:none,stroke:#555,stroke-width:1px
113
-
114
- ```
115
- ---
116
-
117
  ## Quick Start
118
 
119
  ### Docker (recommended, matches the Stage 3 reproduction environment exactly)
 
35
  ![App Architecture](architecture.png)
36
 
37
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
38
  ## Quick Start
39
 
40
  ### Docker (recommended, matches the Stage 3 reproduction environment exactly)