htigenai/reflect_one_4bit

Reflect-One: Fine-tuned Llama 3.1 for Educational Reflection Feedback

This model is a fine-tuned version of Meta's Llama 3.1 8B, specifically optimized for providing thoughtful, constructive feedback on student reflections in interdisciplinary project-based learning environments.

Model Details

• Base Model: unsloth/Meta-Llama-3.1-8B-bnb-4bit
• Training Type: Supervised Fine-Tuning (SFT) using QLoRA
• Context Length: 4096 tokens
• Training Date: March 2025
• License: Llama 3.1 Community License

Intended Use

• Primary Use: Providing constructive feedback on student reflections in interdisciplinary projects
• Intended Users: Educational institutions, instructors, and learning platforms
• Out-of-Scope Use: Any use outside of educational feedback contexts

Training Details

Fine-tuning Approach

• QLoRA fine-tuning with 4-bit quantization
• LoRA rank: 16
• Target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
• Training batch size: 8
• Gradient accumulation steps: 4
• Learning rate: 3e-4
• Training epochs: 3

Training Data

The model was fine-tuned on a curated dataset of student reflections and expert feedback, focusing on:

• Interdisciplinary project work
• Professional identity development
• Stakeholder collaboration
• Problem-solving in ambiguous situations

Performance and Limitations

Capabilities

• Provides structured, constructive feedback aligned with learning outcomes
• Maintains consistent narrative style
• Identifies key learning moments in student reflections
• Suggests areas for deeper reflection

Limitations

• Limited to English language feedback
• Should not be used as the sole source of feedback
• May not capture course-specific technical requirements
• Limited by the context window of 4096 tokens

Ethical Considerations

This model inherits the ethical considerations of the base Llama 3.1 model and adds specific considerations for educational use:

• Should be used as a supplementary tool, not a replacement for human instructors
• Feedback should be reviewed for appropriateness and accuracy
• Model outputs should be transparent to students
• Privacy considerations for student data must be maintained

Versions Available

Two versions of the model are available:

1. htigenai/reflect_one: 16-bit version for higher precision
2. htigenai/reflect_one_4bit: 4-bit quantized version for efficient deployment

Example Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_name = "htigenai/reflect_one"  # or "htigenai/reflect_one_4bit"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    device_map="auto",
    torch_dtype=torch.float16
)

# Format input following Llama 3.1 template
input_text = """<|begin_of_text|><|start_header_id|>system<|end_header_id|>
You are a reflective instructor for ISBEP.<|eot_id|><|start_header_id|>user<|end_header_id|>
[Student reflection here]<|eot_id|><|start_header_id|>assistant<|end_header_id|>"""

# Generate feedback
inputs = tokenizer(input_text, return_tensors="pt").to(model.device)
outputs = model.generate(
    **inputs,
    max_new_tokens=512,
    temperature=0.7,
    do_sample=True
)
response = tokenizer.decode(outputs[0], skip_special_tokens=True)

Citation

If you use this model in your research, please cite:

@misc{reflect_one_2025,
  title={ReflectOne: Fine-tuned Llama 3.1 for Educational Reflection Feedback},
  author={Oliveira, M.J.B., Ruijten - Dodoiu, P.},
  year={2025},
  publisher={unpublished}
}