czszt/llama2-7b-qlora-cover-letter

LLaMA-2-7B-Chat QLoRA Fine-Tuned for Structured Cover Letter Generation

This model is a task-specific fine-tuned version of meta-llama/Llama-2-7b-chat-hf, trained using QLoRA for structured, personalized cover letter generation.
Given a job description and resume fields, the model produces coherent, grounded, and well-structured professional cover letters.

The goal of this project was to improve grounding, reduce hallucinations, and increase template stability compared to the base model.

Note: This repository contains LoRA adapter weights only.
To use the model, load these adapters into the base model
meta-llama/Llama-2-7b-chat-hf, which requires accepting Meta’s license.

---

Model Summary

• Base: LLaMA-2-7B-Chat
• Fine-tuning method: QLoRA (8-bit NF4)
• Task: Cover letter generation from structured job + resume fields
• Dataset: ShashiVish/cover-letter-dataset (cleaned + reformatted)
• Improvement:
  • +65% ROUGE-L vs base
  • +3–5% BERTScore
  • Lower repetition ratio
• Modes: roduces both short (3–6 sentence) and long (multi-paragraph) letters depending on input richness and sampling settings.

---

Model Sources

• Repository: https://github.com/soniatyburczy/llama2-qlora-sft-coverletter-project
  (Training pipeline, evaluation code, and full experimentation logs.)

---

Intended Use

Direct Use

• Automatic cover letter generation
• Structured text generation from job/resume fields
• Template generator for job-application tools

Downstream Use

• Further fine-tuning on HR-specific datasets
• Integration into resume/ATS or recruiting tools
• Chrome extensions (original use case)

Out-of-Scope Use

• Fully automated job applications without human review
• General-purpose long-form writing
• Any high-stakes domain requiring factual accuracy

---

Bias, Risks & Limitations

• Some training samples are low-quality or AI-generated → possible stylistic inconsistency
• Not intended to replace human-written professional communication
• May infer or restate details incorrectly if input fields are sparse
• Performance depends heavily on input structure and clarity

Recommendation: Proofread outputs; keep prompts clean and include key details (“undergraduate,” domain, skills, etc.).

---

How to Use

from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel
import re

base_model = "meta-llama/Llama-2-7b-chat-hf"
adapter_model = "czszt/llama2-coverletter-qlora"

# Load base model + tokenizer
tokenizer = AutoTokenizer.from_pretrained(base_model)
model = AutoModelForCausalLM.from_pretrained(
    base_model,
    device_map="auto",
    torch_dtype="auto"
)

# Load LoRA adapters onto the base model
model = PeftModel.from_pretrained(model, adapter_model)

prompt = """
### Job Description
Job Title: Data Scientist
Company: DataTech Solutions
Preferred Qualifications: SQL, Python, Tableau

### Applicant Resume
Name: John Smith
Current Experience: Data Scientist at DEF Corp
Past Experience: Data Analyst at XYZ Corp
Skills: SQL, Python, Tableau
Qualifications: Bachelor's degree in Statistics

### Cover Letter
Using the information above, write a professional, personalized cover letter.
"""

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(
    **inputs,
    max_new_tokens=512,
    temperature=0.2,
    pad_token_id=tokenizer.eos_token_id  # prevents warnings for LLaMA models
)

output_text = tokenizer.decode(outputs[0], skip_special_tokens=True)

# Optional cleanup: strip everything before "Dear"
text = output_text
if "Dear" in text:
    text = "Dear" + text.split("Dear", 1)[1]

print(text)

---

Training Details

Data

• Dataset: ShashiVish/cover-letter-dataset
• Cleaned: extracted valid cover letter segments
• Prompt schema: structured job + resume fields → cover letter target

Preprocessing

• Regex-based cover letter cleaning
• Removal of “### Sample” artifacts
• Addition of EOS tokens
• 768-token max sequence length

Training Hyperparameters

• Epochs: 4
• Batch size: 1
• Grad accumulation: 4
• Learning rate: 2e-4
• Quantization: 8-bit NF4 double quantization
• LoRA: r=8, alpha=16, target_modules=["q_proj", "v_proj"]
• Precision: FP16
• Optimizer: paged AdamW 32bit

---

Evaluation

Metrics

• ROUGE-L (structure and grounding)
• BERTScore Precision/Recall/F1 (semantic similarity)
• Repetition ratio (hallucination & template stability proxy)

Results (full dataset)

Model	ROUGE-L ↑	BERT-P ↑	BERT-R ↑	BERT-F1 ↑	Repetition ↓
Base	0.314	0.884	0.919	0.901	0.428
Fine-tuned	0.519	0.927	0.939	0.933	0.361

Results (long letters > 500 chars)

Model	ROUGE-L ↑	BERT-P ↑	BERT-R ↑	BERT-F1 ↑	Repetition ↓
Base	0.355	0.904	0.919	0.911	0.429
Fine-tuned	0.507	0.935	0.933	0.934	0.397

---

Limitations

• Dataset contains noisy, unstructured entries
• Model may revert to generic phrasing for weak inputs
• Not human-rated; evaluation relies on automated metrics
• Output may require editing for tone, voice, or specificity

---

Contact

For questions, reach out via GitHub: github.com/soniatyburczy