Upload README.md with huggingface_hub

README.md

@@ -1,190 +1,63 @@
 ---
-language:
-- en
-license: llama3.2
-library_name: transformers
-base_model: meta-llama/Llama-3.2-1B-Instruct
+license: mit
+base_model: meta-llama/Llama-3.2-3B-Instruct
 tags:
 - text-generation
-- llm
 - lora
 - peft
-- fine-tuned
-- creative-writing
 - literature
-- novel
-- storytelling
-- incremental-training
-pipeline_tag: text-generation
-widget:
-- text: Once upon a time, in a distant land,
-  example_title: Story Beginning
-- text: 'Chapter 1: The Beginning
-
-
-    '
-  example_title: Chapter Start
-- text: The old house stood at the edge of the forest,
-  example_title: Scene Setting
-datasets: []
-metrics: []
-model-index:
-- name: NovelCrafter
-  results: []
+- creative-writing
+- dostoevsky
+library_name: peft
 ---
 
-# Model Card: NovelCrafter Fine-Tuned Model
-
-## Model Details
-
-### Model Description
-
-This model is a fine-tuned version of Meta's Llama 3.2 (1B or 3B) using LoRA (Low-Rank Adaptation) on literary text. It has been trained incrementally on book content to capture writing style, narrative patterns, and literary conventions.
-
-- **Developed by**: [990aa](https://github.com/990aa)
-- **Model type**: Causal Language Model (CLM)
-- **Base Model**: 
-  - `meta-llama/Llama-3.2-1B-Instruct` (CPU training)
-  - `meta-llama/Llama-3.2-3B-Instruct` (GPU training)
-- **Language(s)**: English (primarily)
-- **License**: MIT License (training code), Llama 3.2 License (base model)
-- **Finetuned from**: Meta Llama 3.2 Instruct
-- **Training Method**: LoRA (Parameter-Efficient Fine-Tuning)
+# NovelCrafter - Literary LLM
 
-## Uses
+This is a LoRA adapter fine-tuned on literary works using instruction-response pairs.
 
-### Direct Use
-
-This model can be used for:
-- **Text Generation**: Generate text in the style of the training book
-- **Story Continuation**: Continue narratives with consistent style
-- **Creative Writing Assistance**: Help authors write in specific literary styles
-- **Literary Analysis**: Understand patterns in specific works
-- **Educational Purposes**: Learn about fine-tuning and literary AI
-
-### Downstream Use
-
-Can be further fine-tuned on:
-- Additional literary works
-- Specific genres or authors
-- Creative writing tasks
-- Dialogue generation
-- Scene description
-
-### Out-of-Scope Use
-
-This model should NOT be used for:
-- Medical, legal, or financial advice
-- Generating harmful, toxic, or biased content
-- Impersonating specific real individuals
-- Producing academic work without proper attribution
-- Any application requiring factual accuracy without verification
-
-## Bias, Risks, and Limitations
+## Model Details
 
-### Known Limitations
+- **Base Model**: meta-llama/Llama-3.2-3B-Instruct (GPU) / Llama-3.2-1B-Instruct (CPU)
+- **Training Method**: LoRA (Low-Rank Adaptation)
+- **LoRA Rank**: 8
+- **LoRA Alpha**: 32
+- **Target Modules**: q_proj, v_proj
 
-1. **Training Data Bias**: The model reflects biases present in the training literature
-2. **Factual Accuracy**: Not trained for factual tasks; may generate plausible but incorrect information
-3. **Context Length**: Limited to the base model's context window (~8k tokens for Llama 3.2)
-4. **Style Specificity**: Most effective for generating text similar to training material
-5. **Language**: Primarily trained on English text
+## Training Data
 
-### Risks
+The model was trained incrementally on literary works in JSON format with:
+- `previous_chapter`: Context from previous chapter (optional)
+- `instruction`: Writing prompt
+- `response`: Expected literary output
 
-- **Copyright Concerns**: Generated text may inadvertently reproduce training data
-- **Harmful Content**: Despite instruction tuning, may generate inappropriate content
-- **Over-reliance**: Users should not rely solely on model outputs for critical decisions
-- **Hallucination**: May generate confident but false information
+### Completed Training Files
 
-## Training Details
+- dream_of_ridiculous_man1.json
 
-#### Training Hyperparameters
+## Usage
 
-**LoRA Configuration:**
 ```python
-rank (r) = 8
-lora_alpha = 32
-lora_dropout = 0.05
-target_modules = ["q_proj", "v_proj"]
-bias = "none"
-task_type = "CAUSAL_LM"
+from peft import PeftModel
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+base_model = "meta-llama/Llama-3.2-3B-Instruct"
+adapter = "a-01a/novelCrafter"
+
+model = AutoModelForCausalLM.from_pretrained(base_model)
+model = PeftModel.from_pretrained(model, adapter)
+tokenizer = AutoTokenizer.from_pretrained(adapter)
+
+# Generate text
+messages = [
+    {"role": "system", "content": "You are a creative writing assistant."},
+    {"role": "user", "content": "Write a chapter about..."}
+]
+prompt = tokenizer.apply_chat_template(messages, tokenize=False)
+inputs = tokenizer(prompt, return_tensors="pt")
+outputs = model.generate(**inputs, max_new_tokens=500)
+print(tokenizer.decode(outputs[0]))
 ```
 
-**Training Arguments:**
-```python
-num_train_epochs = 3 (per part)
-per_device_train_batch_size = 1
-gradient_accumulation_steps = 8
-learning_rate = 5e-5
-weight_decay = 0.01
-warmup_steps = 100 (adjusted per part)
-fp16 = True (GPU only)
-optimizer = AdamW
-lr_scheduler = Linear with warmup
-```
-
-#### Training Process
-
-1. **Text Extraction**: PDF → plain text
-2. **Chunking**: Split into 10 parts for incremental training
-3. **Tokenization**: Llama tokenizer with max_length=1024
-4. **LoRA Application**: Add trainable adapters to base model
-5. **Incremental Training**: Train on each part sequentially
-6. **Checkpoint Saving**: Save after each part
-7. **Hub Upload**: Push to Hugging Face after each part
-
-**Trainable Parameters:**
-- Total parameters: ~1.2B (1B model) or ~3.2B (3B model)
-- Trainable parameters: ~2.3M (0.07% of total)
-- LoRA enables efficient training with minimal memory
-
-#### Compute Infrastructure
-
-**Hardware:**
-- CPU training: Any modern CPU with 8GB+ RAM
-- GPU training: NVIDIA GPU with 8GB+ VRAM recommended
-- Tested on: Consumer-grade hardware
-
-**Training Time:**
-- CPU (1B model): ~2-4 hours per part (30-40 hours total)
-- GPU (3B model): ~15-30 minutes per part (3-5 hours total)
-
-## Evaluation
-
-### Testing Data
-
-- 10% of each training part held out for evaluation
-- Evaluated using perplexity on held-out test set
-- Real-time evaluation during training
-
-### Metrics
-
-- **Training Loss**: Cross-entropy loss on training data
-- **Validation Loss**: Cross-entropy loss on test data
-- **Perplexity**: exp(validation_loss)
-
-## Technical Specifications
-
-### Model Architecture
-
-- **Base Architecture**: Llama 3.2 (Transformer decoder)
-- **Attention Type**: Multi-head attention with GQA
-- **Hidden Size**: 2048 (1B) or 3072 (3B)
-- **Num Layers**: 16 (1B) or 28 (3B)
-- **Num Attention Heads**: 32
-- **Vocabulary Size**: 128,256
-- **Position Embeddings**: RoPE (Rotary Position Embedding)
-
-### Fine-Tuning Method
-
-**LoRA (Low-Rank Adaptation):**
-- Adds trainable low-rank matrices to attention layers
-- Freezes original model weights
-- Reduces memory and compute requirements
-- Enables efficient multi-task learning
----
+## License
 
-**Model Card Author**: 990aa  
-**Model Card Date**: October 2025  
-**Model Card Version**: 1.0.0
+MIT License