This LLaMA model was fine-tuned on a custom dataset of Jack Sparrow's dialogue extracted from movie scripts, quotes, and paraphrased pirate-style interactions. It's designed to be used in conversational chat interfaces or roleplay environments.

Developed by: ebartha
Model type: Chat bot
Language(s) (NLP): English
Finetuned from model [optional]: Llama-3.2-3B-Instruct

Uses

Direct Use

This model can be used as:

A conversational AI chatbot for fans of Pirates of the Caribbean
A pirate-themed character in a storytelling game or NPC
A personality layer on top of LLM agents (e.g., LangChain, CrewAI)

Out-of-Scope Use

Not suitable for factual Q&A or critical decision-making
Should not be used to impersonate real people or used in harmful, deceptive, or offensive contexts

Bias, Risks, and Limitations

While the model is designed to be humorous and fictional, it is trained on stylized and potentially stereotypical data from movies. Users should not take responses literally or assume cultural accuracy.

Recommendations

Use in entertainment, educational, or fictional settings. Always clarify that the chatbot is a fictional character, not an actual pirate or historical figure.

How to Get Started with the Model

Use the code below to get started with the model.

[More Information Needed]

Training Details

Training Data

The model was fine-tuned on a custom dataset of Jack Sparrow dialogue: Devwa/jackSparrow. The dataset contains over 600 carefully curated and cleaned lines of dialogue from the Pirates of the Caribbean movies, formatted in a ShareGPT-like structure. Each line is embedded in conversational context with pirate-themed prompts, randomized to improve generalization and make the model more interactive.

The data was preprocessed to remove stage directions, formal/non-character lines, and noise (e.g. web references or citations), with additional injections of characteristic Jack Sparrow phrases.

Training Procedure

The model was fine-tuned using Unsloth's SFTTrainer, which wraps around Hugging Face's transformers and trl. The base model was meta-llama/Llama-3.2-3B-Instruct.

Training was conducted using full parameter fine-tuning with LoRA adapters. Each training sample consisted of a single formatted conversation with a user prompt and Jack Sparrow-style response.

Preprocessing [optional]

The dataset text field was text, with pre-formatted prompt/response structure.

Packed sequences were disabled (packing=False) to retain natural dialogue flow.

Each prompt was padded/truncated to a maximum sequence length of 2048 tokens.

Training Hyperparameters

Training regime: mixed precision (fp16 if bfloat16 unsupported, bf16 otherwise)
Epochs: 1
Batch size: 2 (per device)
Gradient accumulation: 4 steps (effective batch size: 8)
Learning rate: 2e-4
Optimizer: AdamW (8-bit)
Weight decay: 0.01
LR scheduler: Linear
Warmup steps: 5
Seed: 3407
Output dir: outputs/

Speeds, Sizes, Times [optional]

PU = Tesla T4. Max memory = 14.748 GB. 2.635 GB of memory reserved. 1:36:40