Model Card for ShAIkespear/Phi-2_DPO_M3_Base_Alt

A LoRA-finetuned and **Direct Preference Optimization (DPO)**–aligned variant of microsoft/phi-2, specialized for multiple-choice question answering (MCQA) with an emphasis on STEM and general knowledge domains. This model represents the alternative base configuration of the final M3 (balanced-then-DPO) training pipeline from the ShAIkespear project. It preserves full precision for highest fidelity and further fine-tuning, without 8-bit quantization.

---

Model Details

• Developed by: ShAIkespear team
• Shared by: ShAIkespear team
• Model type: Causal LM (Phi-2) with LoRA adapters; DPO-aligned
• Languages: English
• License: MIT
• Finetuned from: microsoft/phi-2

Model Sources

• Repository: 2.8B-Phi-2-LLM-QA
• Report: “ShAIkespear – How to replace TAs: A comprehensive study on letting LLMs answer your questions”

---

Uses

Direct Use

• MCQA and educational Q&A (MMLU, OpenBookQA, ScienceQA).
• Alignment research — comparison between DPO training setups (Base vs. Quantized).
• As a high-fidelity reference checkpoint for quantized and downstream variants.

Out-of-Scope Use

• High-stakes or safety-critical applications (medical, legal, policy).
• Generative tasks outside multiple-choice reasoning.
• Misuse in automated exam solving or confidential data leakage.

---

Bias, Risks, and Limitations

• Domain bias: Stronger on factual MCQA, weaker on advanced reasoning tasks.
• Answer drift: May occasionally produce verbose or follow-up answers without explicit formatting.
• Data source risks: EPFL-derived preferences may encode narrow style biases.

Recommendations

• Maintain the structured prompt format:

  ### Question ...
  ### Explanation ...
  ### Answer:
  

• Keep human supervision in any educational or grading use.

• Prefer this full-precision model for fine-tuning or evaluation; use quantized versions for deployment.

---

How to Get Started

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "ShAIkespear/Phi-2_DPO_M3_Base_Alt"

tok = AutoTokenizer.from_pretrained(model_id, use_fast=True)
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto")

prompt = "### Question: Which element has the chemical symbol 'O'?\n### Explanation: The symbol 'O' represents this essential gas.\n### Answer:"
inputs = tok(prompt, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=15)
print(tok.decode(out[0], skip_special_tokens=True))

---

Training Details

Training Data

• SFT stage: Balanced MCQA mix — MathQA, OpenBookQA, ScienceQA, TAL-SCQ5K, and EPFL question sets.
• DPO stage: Human preference pairs (EPFL exams + public feedback datasets like HelpSteer).
• Schema: Unified “### Question / ### Explanation / ### Answer” format.
• Filtering: ≤512 tokens, balanced sample caps (~20k per dataset).

Training Procedure

• Pipeline: SFT → DPO (M3 configuration).
• LoRA parameters: rank = 16, α = 16, dropout = 0.05.
• Batch sizes: SFT = 4; DPO = 1.
• Learning rates: 1e-5 (public) / 1e-4 (EPFL).
• Scheduler: Cosine with warmup.
• Frameworks: Hugging Face Transformers + TRL + PEFT (LoRA).

---

Evaluation Summary

• Configuration: M3 Base (Alt) is the unquantized reference model for the quantized 8-bit variant.
• Performance: Balanced dataset improves cross-domain consistency; DPO enhances answer formatting and style alignment.
• Accuracy: Similar to quantized model (~0.61 MMLU avg.), slightly higher on reasoning subtasks.
• Use case: For experimentation, evaluation, or further domain-specific fine-tuning.

---

Technical Specifications

• Architecture: Phi-2 (~2.78B parameters), decoder-only transformer.
• Objective: SFT next-token prediction + DPO preference alignment.
• Precision: Full precision (fp16/bf16).
• Software: Hugging Face Transformers, TRL, PEFT.

---

Glossary

• MCQA: Multiple-Choice Question Answering
• SFT: Supervised Finetuning
• DPO: Direct Preference Optimization
• LoRA: Low-Rank Adaptation
• Alt (Alternative): Internal naming for the alternate full-precision checkpoint variant of M3