Mistral-7B-v0.3 fine-tuned to write Socratic hints for reading-comprehension questions
Given a passage, a question and its answer, this model writes a short Socratic hint that guides a learner toward the answer without revealing it. It is a LoRA adapter for Mistral-7B-v0.3 trained on MultiHint-9K, 9,987 machine-verified, human-calibrated question–answer–hint tuples in English, Italian and Farsi.
From the paper MultiHint-9K: Orchestrating Human and AI Agency for Scalable Pedagogical Hint Generation (AIED 2026). Code: MultiHint-9K.
ROUGE-L on the held-out set: 0.19 → 0.32 (+57% over the base model).
| Base model | Adapter | ROUGE-L (base → fine-tuned) |
|---|---|---|
| Llama-3.1-8B | MultiHint-9K-Llama-3.1-8B | 0.20 → 0.32 (+52%) |
| Qwen2.5-14B | MultiHint-9K-Qwen2.5-14B | 0.23 → 0.31 (+34%) |
| Mistral-7B-v0.3 | MultiHint-9K-Mistral-7B-v0.3 | 0.19 → 0.32 (+57%) |
| Mistral-Nemo-12B | MultiHint-9K-Nemo-12B | 0.15 → 0.31 (+100%) |
| Mistral-Small-24B | MultiHint-9K-Mistral-Small-24B | 0.18 → 0.33 (+71%) |
All improvements are statistically significant (p < 0.001, Bonferroni-corrected Wilcoxon).
import torch
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
base_model = "mistralai/Mistral-7B-Instruct-v0.3"
tokenizer = AutoTokenizer.from_pretrained(base_model)
model = AutoModelForCausalLM.from_pretrained(base_model, torch_dtype=torch.bfloat16, device_map="auto")
model = PeftModel.from_pretrained(model, "Kamyar-zeinalipour/MultiHint-9K-Mistral-7B-v0.3")
prompt = """You are a Recall Assistant. Provide a succinct Socratic hint for the question based on the passage.
Passage: {passage}
Question: {question}
Answer: {answer}
Hint:"""
Training and evaluation scripts are in the repository.
LoRA r = 32, α = 64, dropout 0.1 on the attention and MLP projections; 3 epochs; AdamW, learning rate 1e-4 with cosine decay; bf16, Flash Attention 2.
@inproceedings{zeinalipour2026multihint,
title = {MultiHint-9K: Orchestrating Human and AI Agency for Scalable Pedagogical Hint Generation},
author = {Kamyar Zeinalipour and Amir Sadeghi and Giovanni Angelini and Leonardo Rigutini and Marco Gori and Marco Maggini},
booktitle = {CEUR Workshop Proceedings},
volume = {4248},
year = {2026},
url = {https://ceur-ws.org/Vol-4248/short11.pdf}
}
Released under the base model's license (apache-2.0).
mistralai/Mistral-7B-Instruct-v0.3