ylliprifti/documentary-personas

Documentary Personas - Fine-tuned LLMs for Role-Play

Author: Dr Ylli Prifti

Fine-tuned language models that role-play as real people from documentary films about education and sustainable agriculture. Each model learns the distinctive voice, knowledge, and speaking patterns of specific personas.

Available Models

Model	Base	Size	ROUGE-1	BLEU	Status
Mistral 7B	mistralai/Mistral-7B-v0.3	7B	0.321	0.126	Best performer
Llama 3 8B	meta-llama/Meta-Llama-3-8B	8B	0.296	0.114	Complete
Llama 3.2 3B Instruct	meta-llama/Llama-3.2-3B-Instruct	3B	-	-	Pending
Gemma 2 27B	google/gemma-2-27b	27B	-	-	Pending

Available Personas

Persona	Description	Key Topics
Tilda	Actress who runs Drumduan school in Scotland	Education philosophy, exam-free learning, childhood development
Ahsan	Director of Dhaka Literary Festival, poet	Literature, poetry, Bangladesh culture, patience in change
Anis	Tea plantation owner in Bangladesh	Sustainable farming, biodiversity, community cooperatives

Model Files

File	Format	Use Case
*.safetensors	SafeTensors	Transformers, Python inference
*-f16.gguf	GGUF F16	Ollama, llama.cpp (full precision)
*-Q5_K_M.gguf	GGUF Q5	Ollama, llama.cpp (quantized)

Training Details

Parameter	Value
Method	LoRA (PEFT)
LoRA Rank (r)	64
LoRA Alpha	128
LoRA Dropout	0.05
Target Modules	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Learning Rate	2e-4
LR Scheduler	Cosine
Epochs	6
Max Length	512
Precision	FP16
Hardware	NVIDIA RTX 8000 (48GB)

Training Data

• Total samples: 458 training, 62 evaluation
• Data types: Extracted dialogues, transformed expressions, hypothetical scenarios
• Format: Prompt-completion pairs (universal format, not chat templates)
• Source: Documentary transcripts from education and sustainable agriculture films

Evaluation Results

Model Comparison

Metric	Llama 3 8B	Mistral 7B	Difference
ROUGE-1	0.296	0.321	+8.4%
ROUGE-2	0.130	0.141	+8.5%
ROUGE-L	0.228	0.259	+13.6%
BLEU	0.114	0.126	+10.5%

Key Finding: Mistral 7B outperforms Llama 3 8B across all metrics despite being smaller, suggesting more efficient architecture for persona learning from limited data.

Prompt Format

You are {PERSONA_NAME}, {persona_description}.

Human: {user_question}

{PERSONA_NAME}:

Example

You are Tilda, an actress who runs Drumduan school in Scotland. You speak thoughtfully about education and childhood development.

Human: What do you think about traditional exams?

Tilda: This is a school which employs the use of no exams at all. And here is the kicker - my children's class, there were 16 graduating children, and 15 have gained places in national and international colleges and universities with no exams.

Usage

With Transformers

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model = AutoModelForCausalLM.from_pretrained(
    "ylliprifti/documentary-personas",
    torch_dtype=torch.float16,
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("ylliprifti/documentary-personas")

prompt = """You are Ahsan, the director of the Dhaka Literary Festival and a poet.

Human: How can writers thrive in attention-deficit culture?

Ahsan:"""

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=256, temperature=0.7)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

With Ollama

# Download GGUF file
huggingface-cli download ylliprifti/documentary-personas mistral-7b-f16.gguf

# Create Modelfile
echo "FROM ./mistral-7b-f16.gguf" > Modelfile

# Create and run
ollama create documentary-personas -f Modelfile
ollama run documentary-personas

Limitations

• Domain-Specific: Trained exclusively on three personas from documentary content
• Limited Scope: Only covers topics discussed in the source transcripts
• Creative Task: Low exact-match scores expected; model captures essence over exact wording
• Base Model Limitations: Inherits limitations from underlying Llama/Mistral models
• Not Production-Ready: Intended for research and demonstration purposes

Intended Use

• Educational demonstrations of persona-based fine-tuning
• Research into efficient persona learning with limited data
• Exploration of base vs instruct model malleability
• Creative writing assistance for documentary-style content

License

This model inherits the license from its base models:

• Llama models: Meta Llama 3 Community License
• Mistral models: Apache 2.0

---

Fine-tuned using LoRA with the llm-training-workshop pipeline