agk4444/sat-tutor-qwen3-8b-gguf

SAT Tutor - Qwen3-8B (GGUF Q4_K_M)

by AGK FIRE INC

A Qwen3-8B fine-tune that tutors for the SAT (math + reading/writing), packaged as a Q4_K_M GGUF for local inference with llama.cpp, Ollama, LM Studio, and friends.

• File: sat-tutor-qwen3-8b.Q4_K_M.gguf (~4.9 GB)
• Format: GGUF V3
• Base model: Qwen/Qwen3-8B
• Adapter (pre-merge): agk4444/sat-tutor-qwen3-8b

What it does

Works through SAT questions step by step and finishes with a clear Answer: X line. Trained on SAT-style math and reading/writing questions with step-by-step solutions.

Held-out eval: 160 questions, identical prompt and decoding for both (greedy, 400 tokens, thinking off), strict scoring = unparseable counted wrong. Reports: eval_base.json, eval_tuned.json.

• Math: base 57% (57/100) → tuned 71% (71/100); tuned answered 97/100 vs base 71/100
• Reading/Writing: base 76.7% (46/60) → tuned 66.7% (40/60) — a real regression; the adapter overfit toward math and wants more RW data before the teacher is frozen
• Overall: base 64.4% (103/160) → tuned 69.4% (111/160)

The tuned adapter was chosen as the quantization target. Q4_K_M on the same suite: 66.9% strict (107/160) — math 67/90 answered (74.4%), reading/writing 40/60 (66.7%), vs tuned fp16 69.4%. Quantization tax ≈ 2.5 pts, all on math; reading/writing unchanged. Report: eval_gguf_q4km.json.

"Parseable" counts questions where the model emitted the expected Answer: X format; "strict" counts unparseable answers as wrong. Unparseable usually means a formatting miss rather than a wrong answer.

Usage

llama.cpp

./llama-cli -m sat-tutor-qwen3-8b.Q4_K_M.gguf -n 512 \
  -p "Solve step by step. If x + 5 = 12, what is x?"

Ollama

# Modelfile
FROM ./sat-tutor-qwen3-8b.Q4_K_M.gguf
TEMPLATE """{{ if .System }}<|im_start|>system
{{ .System }}<|im_end|>{{ end }}{{ if .Prompt }}<|im_start|>user
{{ .Prompt }}<|im_end|>{{ end }}<|im_start|>assistant
"""
PARAMETER stop "<|im_end|>"

ollama create sat-tutor -f Modelfile
ollama run sat-tutor

Running Ollama in Docker

docker run -d --name ollama -p 11434:11434 -v ollama:/root/.ollama ollama/ollama
docker cp sat-tutor-qwen3-8b.Q4_K_M.gguf ollama:/root/
docker cp Modelfile ollama:/root/
docker exec -w /root ollama ollama create sat-tutor -f Modelfile
docker exec ollama ollama run sat-tutor

LM Studio / GPT4All / text-generation-webui

Drop the .gguf into the app's models folder - it auto-detects the Qwen3 chat template.

Prompt format

ChatML (Qwen3-Instruct):

<|im_start|>system
You are a helpful SAT tutor.<|im_end|>
<|im_start|>user
Solve step by step. If x + 5 = 12, what is x?<|im_end|>
<|im_start|>assistant

The model is trained to reason step by step and end with Answer: X.

Training details

• Base: Qwen3-8B, 4-bit QLoRA
• Data: 20k SAT-style examples per run (~12k math incl. MetaMathQA chain-of-thought, ~8k reading/writing incl. RACE), seq len 2048
• Adapter merged into base weights (fp16), then quantized: fp16 -> Q8_0 -> Q4_K_M

Limitations

• SAT-focused: off-topic questions get base-model-quality answers at best.
• 4-bit quantization trades a little accuracy for size; the fp16 merge is the full-quality reference.
• It can still make mistakes - double-check the math against the steps shown.

License

© 2026 AGK FIRE INC. Released under Apache 2.0. (inherited from Qwen3).