1# Example: Gertrude Stein Style Transfer
2 
3A complete case study of training Qwen3-8B-Base to write in Gertrude Stein's style using her 1909 novel "Three Lives."
4 
5## Project Summary
6 
7| Metric | Value |
8|--------|-------|
9| Source Book | Three Lives (1909) |
10| Book Word Count | ~86,000 |
11| Training Examples | 592 |
12| Test Examples | 50 |
13| Model | Qwen/Qwen3-8B-Base |
14| LoRA Rank | 32 |
15| Adapter Size | 352 MB |
16| Training Time | ~15 minutes |
17| Total Cost | ~$2 |
18 
19## Training Metrics
20 
21### Loss Trajectory
22 
23| Step | Train Loss | Test Loss |
24|------|-----------|-----------|
25| 5 | 7579.67 | - |
26| 40 | 4032.35 | 7584.85 |
27| 100 | 2392.17 | 2127.44 |
28| 140 | 2863.17 | 752.95 |
29| 180 | 1327.41 | 229.88 |
30| 200 | 1044.49 | 213.36 |
31| Final | ~800 | ~200 |
32 
33Loss reduction: 97% (7584 to 213 test loss)
34 
35## Style Markers Learned
36 
37The model successfully learned these Gertrude Stein characteristics:
38 
391. **Repetitive sentence structures**: "She was a good woman. She was always a good woman."
402. **Simple vocabulary**: Common words, no ornate language
413. **Comma-separated adjectives**: "a dark, sweet, little, pretty girl"
424. **Present continuous tense**: "She was always doing", "He was thinking"
435. **Character focus**: Deep psychological observation over plot
446. **Rhythmic, hypnotic quality**: Almost musical prose
45 
46## Sample Outputs
47 
48### Modern Scenario: Real Estate Office
49 
50**Prompt**: Write in Gertrude Stein's style about a real estate clerk coming home tired.
51 
52**Output**:
53> It was a very busy day for the clerk in the real estate office. He came home to his small house in the working class part of the town, very tired. He went into the kitchen where his wife was cooking.
54>
55> "Hi, Sis, home for dinner?" he said to her.
56>
57> She looked at him and saw that he was very tired. She looked at him and then looked away into the fire. She said nothing in answer to him.
58>
59> He took off his hat and coat and then sat down on the steps just inside the door to take a rest. He was very tired today, he sat there and just sort of dozed a little.
60 
61**Analysis**: Original content verified - "real estate office", "working class", "Hi, Sis" do not appear in training data.
62 
63### Modern Scenario: Text Messages
64 
65**Prompt**: Write about two lovers communicating through text messages.
66 
67**Output**:
68> "Hey, Herman, I just got a minute to be gone from my patient. Want to talk to you for a little while, Herman," Lena wrote one night, and then she sent the text to him.
69>
70> Herman read it and then he wrote her back, "I was just getting ready to go to bed, Lena. Want me to come and see you tomorrow, Lena. Want me to bring something to you, Lena. Want me to bring you some ice cream, Lena. Want me to bring you some of those little donuts..."
71 
72**Analysis**: Classic Stein anaphora (repetition at sentence starts) applied to modern SMS format.
73 
74## AI Detector Results
75 
76Tested with Pangram AI detector: **100% Human Written**
77 
78Multiple samples tested, all scored as human-written prose.
79 
80## Validation Method
81 
82### Modern Scenario Testing
83 
84We tested the model on scenarios that couldn't exist in 1909:
85- Barista making lattes
86- Social media scrolling
87- Video calls
88- Food delivery drivers
89- Climate change anxiety
90 
91When style markers appeared in modern contexts, it proved the model learned **style** rather than **content**.
92 
93### Originality Verification
94 
95Searched training data for output phrases:
96 
97```bash
98grep "real estate office" dataset.jsonl    # No matches
99grep "working class" dataset.jsonl          # No matches
100grep "Hi, Sis" dataset.jsonl                # No matches
101grep "text messages" dataset.jsonl          # No matches
102```
103 
104## Known Limitations
105 
106### Character Name Leakage (~30% of outputs)
107 
108The model sometimes uses original character names (Melanctha, Mrs. Lehntman, Anna) even in modern scenarios. This is because 592 examples from one book means these names appear hundreds of times.
109 
110**Mitigation**: Train on multiple books by the same author, or add synthetic examples with different names.
111 
112### Success Rate Distribution
113 
114- Perfect style transfer: ~50%
115- Style with name leakage: ~30%
116- Partial style: ~15%
117- Failed: ~5%
118 
119The 50% perfect rate is realistic for an 8B model trained on one book.
120 
121## Configuration Used
122 
123### Dataset Generation
124 
125```python
126CONFIG = {
127    "min_words": 150,
128    "max_words": 400,
129    "overlap": True,  # Last paragraph carried to next chunk
130    "variants_per_chunk": 2,
131    "prompt_templates": 15,
132    "system_prompts": 5,
133    "instruction_model": "gemini-2.0-flash-lite",
134}
135```
136 
137### Training
138 
139```python
140CONFIG = {
141    "model_name": "Qwen/Qwen3-8B-Base",
142    "lora_rank": 32,
143    "learning_rate": 5e-4,
144    "batch_size": 4,
145    "epochs": 3,
146    "eval_every": 20,
147    "save_every": 50,
148}
149```
150 
151## Key Learnings
152 
1531. **Smaller chunks work better**: 150-400 words produced more examples and better style transfer than 250-650
154 
1552. **Prompt diversity is critical**: 15 templates × 5 system prompts = 75 variations prevented memorization
156 
1573. **Base models over instruct**: Qwen3-8B-Base was more malleable than instruct versions
158 
1594. **Modern scenario testing proves transfer**: If style applies to modern contexts, the model learned patterns, not content
160 
1615. **~$2 is enough**: LLM calls for instruction generation (~$0.50) plus Tinker training (~$1.50)
162 
163## Files
164 
165- `sample_outputs.md` - Full model outputs with analysis
166- `training_config.json` - Exact configuration used
167- `dataset_sample.jsonl` - Sample training examples
168 
169