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
• 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)

Model Sources

• Repository: https://github.com/990aa/novelCrafter
• Model Hub: https://huggingface.co/a-01a/novelCrafter

Uses

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

Known Limitations

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

Risks

• 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

Recommendations

Users should:

• Review and edit all generated content
• Add appropriate disclaimers for AI-generated text
• Not use for high-stakes decisions without human oversight
• Be aware of potential copyright issues
• Test thoroughly for their specific use case

Training Details

Training Data

• Source: PDF book(s) placed in input/ directory
• Preprocessing:
  • Text extracted from PDF
  • Cleaned and normalized (whitespace, newlines)
  • Split into sentence chunks (10 sentences per chunk by default)
  • Tokenized with Llama tokenizer
  • 90/10 train/test split per training part

Training Procedure

Training Hyperparameters

LoRA Configuration:

rank (r) = 8
lora_alpha = 32
lora_dropout = 0.05
target_modules = ["q_proj", "v_proj"]
bias = "none"
task_type = "CAUSAL_LM"

Training Arguments:

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

Software:

Python 3.8+
PyTorch 2.0+
Transformers 4.56+
PEFT 0.17+

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)

Note: Specific metrics depend on the training run and can be viewed in WandB logs or training outputs.

Environmental Impact

• Hardware Type: CPU or GPU (varies by user)
• Hours Used: 3-40 hours (depending on hardware)
• Cloud Provider: N/A (local training)
• Compute Region: User-dependent
• Carbon Emitted: Varies by location and power source

I encourage users to:

• Use energy-efficient hardware when possible
• Train during off-peak hours
• Consider renewable energy sources
• Reuse and share trained models

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

Model Card Contact

For questions or concerns about this model:

• GitHub Issues: https://github.com/990aa/novelCrafter/issues
• Email: Via GitHub profile

Changelog

Version 1.0.0 (October 2025)

• Initial release
• Incremental training on literary works
• LoRA fine-tuning implementation
• CPU/GPU optimization
• Hugging Face integration

---

Model Card Authors: 990aa
Model Card Date: October 2025
Model Card Version: 1.0.0