Upload patient_files/windy-pair-en-ilo.md with huggingface_hub
Browse files
patient_files/windy-pair-en-ilo.md
ADDED
|
@@ -0,0 +1,201 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# πͺοΈ Patient File: windy-pair-en-ilo
|
| 2 |
+
**Generated:** 23 Mar 2026 05:49 UTC
|
| 3 |
+
**Pipeline:** Windy Pro Assembly Line Phase 1
|
| 4 |
+
**Built by:** Kit 0C1 Alpha on Veron-1 (RTX 5090, Mount Pleasant SC)
|
| 5 |
+
|
| 6 |
+
---
|
| 7 |
+
|
| 8 |
+
## π Model Information
|
| 9 |
+
|
| 10 |
+
- **Model Key:** `en-ilo`
|
| 11 |
+
- **Model ID:** `windy-pair-en-ilo`
|
| 12 |
+
- **Source Repo:** Helsinki-NLP/opus-mt-en-ilo
|
| 13 |
+
- **Origin:** N/A
|
| 14 |
+
- **License:** CC-BY-4.0
|
| 15 |
+
- **Architecture:** MarianMT (Seq2Seq Transformer)
|
| 16 |
+
|
| 17 |
+
## π Language Pair
|
| 18 |
+
|
| 19 |
+
- **Claimed Direction:** English (`en`) β ILO (`ilo`)
|
| 20 |
+
- **Detected Source Language:** English (`en`)
|
| 21 |
+
|
| 22 |
+
## π
Timeline
|
| 23 |
+
|
| 24 |
+
- **Source Downloaded:** 2026-03-21T02:32:45.514315+00:00
|
| 25 |
+
- **Sweep 1 Certification:**
|
| 26 |
+
- Base: 2026-03-22T19:30:46.884810+00:00
|
| 27 |
+
- CT2: 2026-03-22T19:30:46.884810+00:00
|
| 28 |
+
- LoRA: 2026-03-22T19:30:46.884810+00:00
|
| 29 |
+
- **Sweep 2 Re-certification:** 2026-03-23T01:22:23.957407+00:00
|
| 30 |
+
|
| 31 |
+
## π Re-Certification History
|
| 32 |
+
|
| 33 |
+
- **Total Certification Attempts:** 2
|
| 34 |
+
- **Status Consistent:** CERTIFIED across both sweeps
|
| 35 |
+
|
| 36 |
+
## π¬ Surgery Report β LoRA Variant
|
| 37 |
+
|
| 38 |
+
### LoRA Configuration (from assembly_line.py)
|
| 39 |
+
|
| 40 |
+
```python
|
| 41 |
+
LoraConfig(
|
| 42 |
+
r=4, # LoRA rank
|
| 43 |
+
lora_alpha=8, # Alpha parameter
|
| 44 |
+
target_modules=["q_proj", "v_proj"], # Attention projections
|
| 45 |
+
lora_dropout=0.05,
|
| 46 |
+
bias="none"
|
| 47 |
+
)
|
| 48 |
+
```
|
| 49 |
+
|
| 50 |
+
### Weight Modification Analysis
|
| 51 |
+
|
| 52 |
+
- **Total Model Parameters:** 132,785,664
|
| 53 |
+
- **LoRA Modified Parameters:** 147,456
|
| 54 |
+
- **Percentage Changed:** 0.111%
|
| 55 |
+
- **Noise Factor:** 1e-4 (random perturbation applied to LoRA weights)
|
| 56 |
+
|
| 57 |
+
**LoRA Merge Status:** Merged back into full model (not separate adapters)
|
| 58 |
+
|
| 59 |
+
### Model Size Comparison
|
| 60 |
+
|
| 61 |
+
- **Base Model:** 284.5 MB
|
| 62 |
+
- **LoRA Model:** 284.5 MB (+0.0% vs base)
|
| 63 |
+
- **CT2/INT8 Model:** 188.8 MB (-33.6% vs base)
|
| 64 |
+
|
| 65 |
+
## π Overall Status
|
| 66 |
+
|
| 67 |
+
- **Status:** β
CERTIFIED
|
| 68 |
+
- **Quality Rating:** {'stars': 5.0, 'label': 'Gold Standard', 'avg_cert_score': 10.0, 'best_variant_score': 10, 'variants_tested': 3}
|
| 69 |
+
|
| 70 |
+
## π¬ Sweep 1 Results (Initial Certification)
|
| 71 |
+
|
| 72 |
+
| Variant | Status | Score | Stars | Quality | HF Repo |
|
| 73 |
+
|---------|--------|-------|-------|---------|----------|
|
| 74 |
+
| Base | β
CERTIFIED | 10/10 | βββββ 5.0 | Gold Standard | WindyLabs/windy-pair-en-ilo |
|
| 75 |
+
| CT2/INT8 | β
CERTIFIED | 10/10 | βββββ 5.0 | Gold Standard | WindyLabs/windy-pair-en-ilo-ct2 |
|
| 76 |
+
| LoRA | β
CERTIFIED | 10/10 | βββββ 5.0 | Gold Standard | WindyLabs/windy-pair-en-ilo-lora |
|
| 77 |
+
|
| 78 |
+
### Sample Outputs (Sweep 1)
|
| 79 |
+
|
| 80 |
+
**Base Variant:**
|
| 81 |
+
|
| 82 |
+
1. β
Input: `Hello, how are you today?`
|
| 83 |
+
Output: `Hello, ania ti kasasaadmo ita?`
|
| 84 |
+
|
| 85 |
+
2. β
Input: `The weather is beautiful this morning.`
|
| 86 |
+
Output: `Nagpintas ti paniempo ita nga agsapa.`
|
| 87 |
+
|
| 88 |
+
3. β
Input: `I would like to order a cup of coffee, please.`
|
| 89 |
+
Output: `Kayatko ti agorder iti kape, pangngaasim.`
|
| 90 |
+
|
| 91 |
+
**CT2/INT8 Variant:**
|
| 92 |
+
|
| 93 |
+
1. β
Input: `Hello, how are you today?`
|
| 94 |
+
Output: `Hello, komustaka ita?`
|
| 95 |
+
|
| 96 |
+
2. β
Input: `The weather is beautiful this morning.`
|
| 97 |
+
Output: `Nagpintas ti paniempo ita nga agsapa.`
|
| 98 |
+
|
| 99 |
+
3. β
Input: `I would like to order a cup of coffee, please.`
|
| 100 |
+
Output: `Kayatko ti agorder iti sangatasa a kape, pangngaasiyo.`
|
| 101 |
+
|
| 102 |
+
**LoRA Variant:**
|
| 103 |
+
|
| 104 |
+
1. β
Input: `Hello, how are you today?`
|
| 105 |
+
Output: `Hello, ania ti kasasaadmo ita?`
|
| 106 |
+
|
| 107 |
+
2. β
Input: `The weather is beautiful this morning.`
|
| 108 |
+
Output: `Nagpintas ti paniempo ita nga agsapa.`
|
| 109 |
+
|
| 110 |
+
3. β
Input: `I would like to order a cup of coffee, please.`
|
| 111 |
+
Output: `Kayatko ti agorder iti kape, pangngaasim.`
|
| 112 |
+
|
| 113 |
+
## π¬ Sweep 2 Results (Re-certification with Correct Source Language)
|
| 114 |
+
|
| 115 |
+
- **Status:** β
CERTIFIED
|
| 116 |
+
- **Date:** 2026-03-23T01:22:23.957407+00:00
|
| 117 |
+
|
| 118 |
+
| Variant | Certified | Score | Stars | Quality |
|
| 119 |
+
|---------|-----------|-------|-------|----------|
|
| 120 |
+
| Base | β
True | 10/10 | βββββ 5.0 | Premium |
|
| 121 |
+
| CT2/INT8 | β
True | 10/10 | βββββ 5.0 | Premium |
|
| 122 |
+
| LoRA | β
True | 10/10 | βββββ 5.0 | Premium |
|
| 123 |
+
|
| 124 |
+
### Sample Outputs (Sweep 2)
|
| 125 |
+
|
| 126 |
+
**Base Variant:**
|
| 127 |
+
|
| 128 |
+
1. β
Input: `Hello, how are you today?`
|
| 129 |
+
Output: `Hello, ania ti kasasaadmo ita?`
|
| 130 |
+
|
| 131 |
+
2. β
Input: `The weather is beautiful this morning.`
|
| 132 |
+
Output: `Nagpintas ti paniempo ita nga agsapa.`
|
| 133 |
+
|
| 134 |
+
3. β
Input: `I would like to order a cup of coffee, please.`
|
| 135 |
+
Output: `Kayatko ti agorder iti kape, pangngaasim.`
|
| 136 |
+
|
| 137 |
+
**CT2/INT8 Variant:**
|
| 138 |
+
|
| 139 |
+
1. β
Input: `Hello, how are you today?`
|
| 140 |
+
Output: `Hello, komustaka ita?`
|
| 141 |
+
|
| 142 |
+
2. β
Input: `The weather is beautiful this morning.`
|
| 143 |
+
Output: `Nagpintas ti paniempo ita nga agsapa.`
|
| 144 |
+
|
| 145 |
+
3. β
Input: `I would like to order a cup of coffee, please.`
|
| 146 |
+
Output: `Kayatko ti agorder iti sangatasa a kape, pangngaasiyo.`
|
| 147 |
+
|
| 148 |
+
**LoRA Variant:**
|
| 149 |
+
|
| 150 |
+
1. β
Input: `Hello, how are you today?`
|
| 151 |
+
Output: `Hello, ania ti kasasaadmo ita?`
|
| 152 |
+
|
| 153 |
+
2. β
Input: `The weather is beautiful this morning.`
|
| 154 |
+
Output: `Nagpintas ti paniempo ita nga agsapa.`
|
| 155 |
+
|
| 156 |
+
3. β
Input: `I would like to order a cup of coffee, please.`
|
| 157 |
+
Output: `Kayatko ti agorder iti kape, pangngaasim.`
|
| 158 |
+
|
| 159 |
+
## π©Ί Symptoms
|
| 160 |
+
|
| 161 |
+
- β
No issues detected
|
| 162 |
+
|
| 163 |
+
## π‘ Hypothesis / Analysis
|
| 164 |
+
|
| 165 |
+
- β
Model successfully certified - all variants meet quality thresholds
|
| 166 |
+
|
| 167 |
+
## ποΈ Architecture Details
|
| 168 |
+
|
| 169 |
+
- **Model Type:** MarianMT (Seq2Seq Transformer)
|
| 170 |
+
- **Hidden Size (d_model):** 512
|
| 171 |
+
- **Encoder Layers:** 6
|
| 172 |
+
- **Decoder Layers:** 6
|
| 173 |
+
- **Vocab Size:** 57,777
|
| 174 |
+
- **Attention Heads:** 8
|
| 175 |
+
|
| 176 |
+
---
|
| 177 |
+
*Patient file generated by Windy Pro Patient File Generator v3.0 (Admiral Edition)*
|
| 178 |
+
|
| 179 |
+
---
|
| 180 |
+
|
| 181 |
+
## OPUS-100 Deep Fine-Tune (Herm Zero, Dr. B)
|
| 182 |
+
- **Date:** 2026-03-26T13:58:28 UTC
|
| 183 |
+
- **Doctor:** Herm Zero (Dr. B) β Herm 0, First Claude Code, Kit Army Fleet
|
| 184 |
+
- **Machine:** Veron-1 (RTX 5090, Mount Pleasant SC)
|
| 185 |
+
- **Result:** IMPROVED
|
| 186 |
+
- **Base Score:** 86.1/100 (4.5 stars)
|
| 187 |
+
- **Improved Score:** 93.2/100 (4.5 stars)
|
| 188 |
+
- **Score Improvement:** +7.1 points
|
| 189 |
+
- **Training Data:** 2,093 samples from OPUS-100/Tatoeba/WikiMatrix
|
| 190 |
+
- **Method:** Full weight fine-tune (lr=1e-5, 1 epoch, mixed precision fp16)
|
| 191 |
+
- **Weights:** herm0/model.safetensors
|
| 192 |
+
- **CT2 Updated:** Yes (herm0 weights propagated to ct2/ directory)
|
| 193 |
+
|
| 194 |
+
---
|
| 195 |
+
|
| 196 |
+
## CT2 Safetensors Re-Export (Herm Zero, Dr. B)
|
| 197 |
+
- **Date:** 2026-03-24 ~15:18 UTC
|
| 198 |
+
- **Doctor:** Herm Zero (Dr. B)
|
| 199 |
+
- **Procedure:** Fixed broken pickle INT8 format β re-exported as proper safetensors
|
| 200 |
+
- **Reason:** transformers 4.50+ broke INT8 pickle loader compatibility
|
| 201 |
+
- **Method:** Load base model via MarianMTModel.from_pretrained(), save_pretrained() to ct2/
|