Upload DATA_FORMAT.md with huggingface_hub
Browse files- DATA_FORMAT.md +34 -0
DATA_FORMAT.md
CHANGED
|
@@ -52,6 +52,26 @@ Prediction quality is better if your history CSV also includes:
|
|
| 52 |
- `away_tactic_id`
|
| 53 |
- `home_coach_id`
|
| 54 |
- `away_coach_id`
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 55 |
|
| 56 |
These optional fields help the wrapper compute richer features with fewer fallbacks.
|
| 57 |
|
|
@@ -65,7 +85,9 @@ match_id,date,home_team,away_team,home_goals,away_goals,home_team_id,away_team_i
|
|
| 65 |
The bundled synthetic history currently includes:
|
| 66 |
- `120` rows
|
| 67 |
- `20` La Liga team names
|
|
|
|
| 68 |
- enough depth for rolling-form demo calculations
|
|
|
|
| 69 |
|
| 70 |
## Fixture CSV Requirements
|
| 71 |
For `predict_batch.py`, the fixtures CSV should contain:
|
|
@@ -115,6 +137,18 @@ If you keep using only the synthetic sample files:
|
|
| 115 |
- the methods will behave correctly
|
| 116 |
- but the predictions should be treated as demonstration-quality, not production-quality
|
| 117 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 118 |
## What Not To Publish
|
| 119 |
Do not publish large internal historical CSVs unless redistribution rights are explicit.
|
| 120 |
|
|
|
|
| 52 |
- `away_tactic_id`
|
| 53 |
- `home_coach_id`
|
| 54 |
- `away_coach_id`
|
| 55 |
+
- `home_player_minutes_total_prev5`
|
| 56 |
+
- `away_player_minutes_total_prev5`
|
| 57 |
+
- `home_player_goals_total_prev5`
|
| 58 |
+
- `away_player_goals_total_prev5`
|
| 59 |
+
- `home_player_assists_total_prev5`
|
| 60 |
+
- `away_player_assists_total_prev5`
|
| 61 |
+
- `home_player_yellow_cards_total_prev5`
|
| 62 |
+
- `away_player_yellow_cards_total_prev5`
|
| 63 |
+
- `home_player_red_cards_total_prev5`
|
| 64 |
+
- `away_player_red_cards_total_prev5`
|
| 65 |
+
- `home_player_starters_count_prev5`
|
| 66 |
+
- `away_player_starters_count_prev5`
|
| 67 |
+
- `home_player_used_count_prev5`
|
| 68 |
+
- `away_player_used_count_prev5`
|
| 69 |
+
- `home_player_injured_count_prev5`
|
| 70 |
+
- `away_player_injured_count_prev5`
|
| 71 |
+
- `home_player_suspended_count_prev5`
|
| 72 |
+
- `away_player_suspended_count_prev5`
|
| 73 |
+
- `home_tactic_stability_last5`
|
| 74 |
+
- `away_tactic_stability_last5`
|
| 75 |
|
| 76 |
These optional fields help the wrapper compute richer features with fewer fallbacks.
|
| 77 |
|
|
|
|
| 85 |
The bundled synthetic history currently includes:
|
| 86 |
- `120` rows
|
| 87 |
- `20` La Liga team names
|
| 88 |
+
- `35` columns
|
| 89 |
- enough depth for rolling-form demo calculations
|
| 90 |
+
- richer optional context columns for better demonstration quality
|
| 91 |
|
| 92 |
## Fixture CSV Requirements
|
| 93 |
For `predict_batch.py`, the fixtures CSV should contain:
|
|
|
|
| 137 |
- the methods will behave correctly
|
| 138 |
- but the predictions should be treated as demonstration-quality, not production-quality
|
| 139 |
|
| 140 |
+
## Why The Model Can Use 48 Signals Even If The CSV Has Fewer Columns
|
| 141 |
+
The trained model expects `48` numeric input signals at inference time.
|
| 142 |
+
|
| 143 |
+
Those signals are produced in three ways:
|
| 144 |
+
1. directly from columns already present in the history CSV
|
| 145 |
+
2. derived by the wrapper from past match rows
|
| 146 |
+
3. filled by fallback defaults when richer optional context is missing
|
| 147 |
+
|
| 148 |
+
So the public rule is:
|
| 149 |
+
- the CSV does not need to contain all `48` final model inputs as raw columns
|
| 150 |
+
- but a richer CSV allows the wrapper to produce a stronger feature row
|
| 151 |
+
|
| 152 |
## What Not To Publish
|
| 153 |
Do not publish large internal historical CSVs unless redistribution rights are explicit.
|
| 154 |
|