agk4444 / sat-tutor-qwen3-8b

sat-tutor-qwen3-8b

Developed by AGK FIRE INC .

A LoRA adapter on Qwen/Qwen3-8B fine-tuned for SAT tutoring — step-by-step math problem solving and Reading & Writing (evidence-based) practice questions.

Model details

• Base model: Qwen/Qwen3-8B
• Method: 4-bit QLoRA (PEFT 0.21.0)
• LoRA config: rank 16, alpha 32, dropout 0.05 — all linear projections ( q/k/v/o , gate/up/down )
• Trainable parameters: ~43.6M (0.53% of 8.23B total)
• Hardware: Kaggle 2×T4 GPUs
• Developed by: AGK FIRE INC

Intended use

• Direct use: SAT practice tutoring — math and Reading & Writing questions with worked explanations.
• Out of scope: Not a substitute for official SAT prep materials; always verify solutions before relying on them.

Training details

• Examples: 17,000 (1 epoch, 850 steps)
• Mix: 40% Reading & Writing (6,800) · 60% math
• Sources: SAT-style data, RACE (all), MetaMathQA
• Hyperparameters: LR 2e-4, sequence length 2048
• Instruction format: User: / Tutor: turns (see chat template below)

How to use

This is an adapter only — load it on top of the base model:

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

tok = AutoTokenizer.from_pretrained("agk4444/sat-tutor-qwen3-8b")
base = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen3-8B",
    torch_dtype=torch.bfloat16,
    device_map="auto",
)
model = PeftModel.from_pretrained(base, "agk4444/sat-tutor-qwen3-8b")
model.eval()

messages = [
    {"role": "user", "content": "Solve: if 3x + 5 = 20, what is x?"},
]
prompt = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tok(prompt, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=512)
print(tok.decode(out[0], skip_special_tokens=True))

Prompt format (chat template bundled in this repo):

<|im_start|>user
<your question><|im_end|>
<|im_start|>assistant

Evaluation

Benchmarks against the base model are in progress (planned: AGIEval SAT math and Reading & Writing splits). Results will be posted here when available.

Bias, risks, and limitations

• The tutor can produce incorrect math or misread passages — double-check answers against official materials.
• Trained on English SAT-style data; performance outside that domain is untested.
• Inherits the biases and limitations of the Qwen3-8B base model.

Environmental impact

• Hardware: 2× NVIDIA T4 (Kaggle)
• Duration: ~7 wall-clock hours

© 2026 AGK FIRE INC. Released under Apache 2.0.