Model Card for llama2-7b-instruction-tuned (Step 4 Model)

This model is the result of Step 4 of INFO 7374 Assignment 3, implementing the "Self-Alignment with Instruction Backtranslation" paper (https://arxiv.org/pdf/2308.06259.pdf) using LoRA-based finetuning. It was trained on OpenAssistant-Guanaco seed data, followed by synthetic instruction generation from LIMA completions and LLM-based filtering.

Model Details

Model Description

• Developed by: Niki Choksi (for INFO 7374 Assignment 3)
• Model type: Causal Language Model (Instruction-tuned)
• Language(s): English
• License: Apache 2.0 (inherits from base model)
• Finetuned from model: meta-llama/Llama-2-7b-hf

Model Sources

• Paper: https://arxiv.org/pdf/2308.06259.pdf
• Notebook: [Colab Link Here]
• Dataset: openassistant-guanaco, LIMA

Uses

Direct Use

This model can be used for zero-shot or few-shot instruction following tasks. It predicts instructions given outputs during backward model training, and later produces aligned responses for synthetic instruction/response pairs.

Downstream Use

The model can serve as a base for further instruction tuning or evaluation research.

Out-of-Scope Use

• Multi-turn dialog (was explicitly filtered out)
• Non-English generation
• Sensitive or factual use without validation

Bias, Risks, and Limitations

• The model inherits any biases from the LLaMA-2-7B base model and the training datasets
• Instructions and responses are filtered by an LLM, which may introduce or reinforce subjective judgment

Recommendations

• Evaluate outputs manually before downstream use
• Use the included LLM rating logic for additional sample quality filtering

How to Get Started with the Model

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("Nikichoksi/llama2-7b-instruction-tuned")
tokenizer = AutoTokenizer.from_pretrained("Nikichoksi/llama2-7b-instruction-tuned")

prompt = "Write a haiku about the moon."
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=100)
print(tokenizer.decode(outputs[0]))

Training Details

Training Data

• Seed: openassistant-guanaco
• Synthetic: 150 single-turn samples from LIMA completions
• Filtering: Synthetic data was filtered using LLM rating (scale of 1–5)

Training Procedure

Preprocessing

• Filtered out multi-turn examples from LIMA completions
• Applied backtranslation to generate instructions from outputs

Training Hyperparameters

• LoRA rank: 8
• Batch size: 4
• Precision: bf16 (on Colab GPU)
• Optimizer: AdamW

Speeds, Sizes, Times

• Model finetuned for ~2 hours on Colab A100

---

Evaluation

Testing Data, Factors & Metrics

• LLM evaluator: meta-llama/Llama-2-7b-chat-hf
• Metric: 1–5 quality rating of instruction/response pairs
• Goal: Maximize count of samples with rating ≥ 4

Results

• 5 high-quality and 5 low-quality examples logged
• Final model produces coherent, instruction-following completions

---

Environmental Impact

• Hardware Type: Colab A100 GPU
• Hours used: ~2
• Cloud Provider: Google Colab
• Compute Region: US
• Carbon Emitted: Low (<1 kg CO2eq)

---

Technical Specifications

Model Architecture and Objective

• Based on LLaMA 2 7B architecture
• Causal language modeling objective with LoRA adapters

Compute Infrastructure

• Hardware: 1 x A100 (Colab)
• Software:
  • PEFT 0.16.0
  • Transformers 4.39+
  • BitsAndBytes 0.41.1

---

Model Card Contact

• Author: Niki Choksi
• Email: [email protected]