Upload patient_files/windy-pair-eo-cs.md with huggingface_hub

patient_files/windy-pair-eo-cs.md

@@ -0,0 +1,201 @@
+# 🌪️ Patient File: windy-pair-eo-cs
+**Generated:** 23 Mar 2026 05:49 UTC
+**Pipeline:** Windy Pro Assembly Line Phase 1
+**Built by:** Kit 0C1 Alpha on Veron-1 (RTX 5090, Mount Pleasant SC)
+
+---
+
+## 📋 Model Information
+
+- **Model Key:** `eo-cs`
+- **Model ID:** `windy-pair-eo-cs`
+- **Source Repo:** Helsinki-NLP/opus-mt-eo-cs
+- **Origin:** N/A
+- **License:** CC-BY-4.0
+- **Architecture:** MarianMT (Seq2Seq Transformer)
+
+## 🌍 Language Pair
+
+- **Claimed Direction:** Esperanto (`eo`) → Czech (`cs`)
+- **Detected Source Language:** Esperanto (`eo`)
+
+## 📅 Timeline
+
+- **Source Downloaded:** 2026-03-21T03:06:01.207927+00:00
+- **Sweep 1 Certification:**
+  - Base: 2026-03-22T19:32:04.677416+00:00
+  - CT2: 2026-03-22T19:32:04.677416+00:00
+  - LoRA: 2026-03-22T19:32:04.677416+00:00
+- **Sweep 2 Re-certification:** 2026-03-23T01:31:31.493429+00:00
+
+## 🔄 Re-Certification History
+
+- **Total Certification Attempts:** 2
+- **Status Consistent:** CERTIFIED across both sweeps
+
+## 🔬 Surgery Report — LoRA Variant
+
+### LoRA Configuration (from assembly_line.py)
+
+```python
+LoraConfig(
+    r=4,              # LoRA rank
+    lora_alpha=8,     # Alpha parameter
+    target_modules=["q_proj", "v_proj"],  # Attention projections
+    lora_dropout=0.05,
+    bias="none"
+)
+```
+
+### Weight Modification Analysis
+
+- **Total Model Parameters:** 55,401,984
+- **LoRA Modified Parameters:** 147,456
+- **Percentage Changed:** 0.266%
+- **Noise Factor:** 1e-4 (random perturbation applied to LoRA weights)
+
+**LoRA Merge Status:** Merged back into full model (not separate adapters)
+
+### Model Size Comparison
+
+- **Base Model:** 183.6 MB
+- **LoRA Model:** 183.6 MB (+0.0% vs base)
+- **CT2/INT8 Model:** 63.4 MB (-65.5% vs base)
+
+## 📊 Overall Status
+
+- **Status:** ✅ CERTIFIED
+- **Quality Rating:** {'stars': 5.0, 'label': 'Gold Standard', 'avg_cert_score': 10.0, 'best_variant_score': 10, 'variants_tested': 3}
+
+## 🔬 Sweep 1 Results (Initial Certification)
+
+| Variant | Status | Score | Stars | Quality | HF Repo |
+|---------|--------|-------|-------|---------|----------|
+| Base | ✅ CERTIFIED | 8/10 | ⭐⭐⭐⭐ 4.0 | Premium | WindyLabs/windy-pair-eo-cs |
+| CT2/INT8 | ✅ CERTIFIED | 8/10 | ⭐⭐⭐⭐ 4.0 | Premium | WindyLabs/windy-pair-eo-cs-ct2 |
+| LoRA | ✅ CERTIFIED | 8/10 | ⭐⭐⭐⭐ 4.0 | Premium | WindyLabs/windy-pair-eo-cs-lora |
+
+### Sample Outputs (Sweep 1)
+
+**Base Variant:**
+
+1. ❌ Input: `Hello, how are you today?`
+   Output: `Hello, hawy?`
+
+2. ❌ Input: `The weather is beautiful this morning.`
+   Output: `Thomasweather condition.`
+
+3. ✅ Input: `I would like to order a cup of coffee, please.`
+   Output: `Io dluhopisů likvidace copeff, psaní.`
+
+**CT2/INT8 Variant:**
+
+1. ❌ Input: `Hello, how are you today?`
+   Output: `Hello, hawy?`
+
+2. ❌ Input: `The weather is beautiful this morning.`
+   Output: `Thomasweather condition.`
+
+3. ✅ Input: `I would like to order a cup of coffee, please.`
+   Output: `Io dluhopisů likvidace creaffe, psaní.`
+
+**LoRA Variant:**
+
+1. ❌ Input: `Hello, how are you today?`
+   Output: `Hello, hawy?`
+
+2. ❌ Input: `The weather is beautiful this morning.`
+   Output: `Thomasweather condition.`
+
+3. ✅ Input: `I would like to order a cup of coffee, please.`
+   Output: `Io dluhopisů likvidace copeff, psaní.`
+
+## 🔬 Sweep 2 Results (Re-certification with Correct Source Language)
+
+- **Status:** ✅ CERTIFIED
+- **Date:** 2026-03-23T01:31:31.493429+00:00
+
+| Variant | Certified | Score | Stars | Quality |
+|---------|-----------|-------|-------|----------|
+| Base | ✅ True | 10/10 | ⭐⭐⭐⭐⭐ 5.0 | Premium |
+| CT2/INT8 | ✅ True | 10/10 | ⭐⭐⭐⭐⭐ 5.0 | Premium |
+| LoRA | ✅ True | 10/10 | ⭐⭐⭐⭐⭐ 5.0 | Premium |
+
+### Sample Outputs (Sweep 2)
+
+**Base Variant:**
+
+1. ✅ Input: `Bonan tagon, kiel vi fartas hodiaŭ?`
+   Output: `Dobrý den, jak se máš dnes?`
+
+2. ✅ Input: `La infanoj ludas en la parko post la lernejo.`
+   Output: `Děti hrají v parku po škole.`
+
+3. ✅ Input: `Bonvolu helpi min trovi la bibliotekon.`
+   Output: `Prosím, pomoz mi najít knihovnu.`
+
+**CT2/INT8 Variant:**
+
+1. ✅ Input: `Bonan tagon, kiel vi fartas hodiaŭ?`
+   Output: `Dobrý den, jak se máš dnes?`
+
+2. ✅ Input: `La infanoj ludas en la parko post la lernejo.`
+   Output: `Děti hrají v parku po škole.`
+
+3. ✅ Input: `Bonvolu helpi min trovi la bibliotekon.`
+   Output: `Prosím, pomoz mi najít knihovnu.`
+
+**LoRA Variant:**
+
+1. ✅ Input: `Bonan tagon, kiel vi fartas hodiaŭ?`
+   Output: `Dobrý den, jak se máš dnes?`
+
+2. ✅ Input: `La infanoj ludas en la parko post la lernejo.`
+   Output: `Děti hrají v parku po škole.`
+
+3. ✅ Input: `Bonvolu helpi min trovi la bibliotekon.`
+   Output: `Prosím, pomoz mi najít knihovnu.`
+
+## 🩺 Symptoms
+
+- ✅ No issues detected
+
+## 💡 Hypothesis / Analysis
+
+- ✅ Model successfully certified - all variants meet quality thresholds
+
+## 🏗️ Architecture Details
+
+- **Model Type:** MarianMT (Seq2Seq Transformer)
+- **Hidden Size (d_model):** 512
+- **Encoder Layers:** 6
+- **Decoder Layers:** 6
+- **Vocab Size:** 7,397
+- **Attention Heads:** 8
+
+---
+*Patient file generated by Windy Pro Patient File Generator v3.0 (Admiral Edition)*
+
+---
+
+## OPUS-100 Deep Fine-Tune (Herm Zero, Dr. B)
+- **Date:** 2026-03-26T18:39:04 UTC
+- **Doctor:** Herm Zero (Dr. B) — Herm 0, First Claude Code, Kit Army Fleet
+- **Machine:** Veron-1 (RTX 5090, Mount Pleasant SC)
+- **Result:** IMPROVED
+- **Base Score:** 90.9/100 (4.5 stars)
+- **Improved Score:** 91.5/100 (4.5 stars)
+- **Score Improvement:** +0.5 points
+- **Training Data:** 50,000 samples from OPUS-100/Tatoeba/WikiMatrix
+- **Method:** Full weight fine-tune (lr=1e-5, 1 epoch, mixed precision fp16)
+- **Weights:** herm0/model.safetensors
+- **CT2 Updated:** Yes (herm0 weights propagated to ct2/ directory)
+
+---
+
+## CT2 Safetensors Re-Export (Herm Zero, Dr. B)
+- **Date:** 2026-03-24 ~15:18 UTC
+- **Doctor:** Herm Zero (Dr. B)
+- **Procedure:** Fixed broken pickle INT8 format — re-exported as proper safetensors
+- **Reason:** transformers 4.50+ broke INT8 pickle loader compatibility
+- **Method:** Load base model via MarianMTModel.from_pretrained(), save_pretrained() to ct2/