llama-3.2-3b-qlora-finetuned-hr

Model Card for Model ID

<!-- Provide a quick summary of what the model is/does. -->

This model has been trained to recommend Nature of Action (NOA) codes for processing personnel action requests for federal employees based on common scenarios.

Model Details

Model Description

<!-- Provide a longer summary of what this model is. -->

This model has been trained to recommend Nature of Action (NOA) for processing personnel action requests for federal employees based on common scenarios. NOA codes are unique numerical identifiers for every personnel action (like promotions, appointments, or retirements) in the U.S. Federal government, used for data processing on forms like the SF-50, with categories 000-899 for standard actions and 900-999 for agency-specific matters, all detailed in OPM's Guide to Processing Personnel Actions (GPPA).

This is the model card of a 🤗 transformers model that has been pushed on the Hub. This model card has been automatically generated.

• Developed by: Frank Manja
• Model type: question-answering
• Language(s) (NLP): English
• License: Apache 2.0
• Finetuned from model: meta-llama/Llama-3.2-1B

Uses

<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->

This is a demonstration model and not intended for production use. Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model.

QLoRA Fine-tuning of Llama-3.2-3B for HR NOA Codes

This section outlines the fine-tuning process of the meta-llama/Llama-3.2-3B model using QLoRA for generating OPM NOA (Nature of Action) codes related to HR scenarios.

Training Regime

• Fine-tuning Method: QLoRA (Quantized Low-Rank Adaptation)
• Base Model: meta-llama/Llama-3.2-3B

Quantization Details

• load_in_4bit: True
• bnb_4bit_quant_type: nf4 (NormalFloat 4-bit)
• bnb_4bit_use_double_quant: True
• bnb_4bit_compute_dtype: torch.bfloat16

LoRA Configuration Parameters

• r: 8 (LoRA attention dimension)
• lora_alpha: 16 (Alpha parameter for LoRA scaling)
• lora_dropout: 0.05 (Dropout probability for LoRA layers)
• bias: none
• task_type: CAUSAL_LM
• target_modules: ["q_proj", "v_proj", "k_proj", "o_proj", "gate_proj", "up_proj", "down_proj"]

Training Hyperparameters

• output_dir: ./results_llama_3_2_3b_qlora
• num_train_epochs: 1
• per_device_train_batch_size: 4
• gradient_accumulation_steps: 2
• optim: paged_adamw_8bit
• save_steps: 100
• logging_steps: 100
• learning_rate: 2e-4
• weight_decay: 0.001
• fp16: False
• bf16: True
• max_grad_norm: 0.3
• max_steps: -1 (disabled for epoch-based training)
• warmup_ratio: 0.03
• group_by_length: True
• lr_scheduler_type: constant
• report_to: none

Speeds, Sizes, Times

• GPU Used: Tesla T4 (with 15360 MiB VRAM)
• Dataset Size: 50 examples
• Model Size: Llama-3.2-3B (base model)
• Training Epochs: 1
• Training Duration: The exact training duration was not logged, but the process completed successfully within a single epoch on the specified GPU.

Testing Data

No explicit testing data was used for evaluation in this fine-tuning process. The focus was on demonstrating the QLoRA fine-tuning workflow.

Factors

No specific factors beyond the standard model architecture and QLoRA configuration were explicitly controlled or analyzed for their impact on performance during this initial fine-tuning run.

Metrics

No quantitative metrics (e.g., ROUGE, BLEU, perplexity) were used to evaluate the fine-tuned model's performance. Evaluation was qualitative based on generated text.

Results

Below is a qualitative inference example demonstrating the model's output after QLoRA fine-tuning:

Model Card Authors

Frank Manja

Model Card Contact

Frank Manja (fmanja on HuggingFace)

Disclaimer: All views and opinions expressed here are my own and do not represent those of my employer, past or present, or any affiliated organizations. Any work shared is personal and created outside my professional duties.