DrUkachi/ktt-math-tutor-models

KTT Math Tutor — Models

Companion model artefacts for the AIMS KTT Hackathon Tier-3 submission S2.T3.1 AI Math Tutor for Early Learners. Source code and training scripts: https://github.com/DrUkachi/ktt-math-tutor.

What's here

Subfolder / file	Size	Role
whisper-tiny-child-lora-ct2int8/	44 MB	child-voice LoRA-tuned Whisper-tiny, merged, CTranslate2 int8 for CPU
tinyllama-numeracy-qlora-adapter/	21 MB	QLoRA adapter (r=16, NF4 base) trained on 200 synthetic numeracy instructions
tinyllama-numeracy-Q4_K_M.gguf	637 MB	the adapter merged into TinyLlama-1.1B and quantised to Q4_K_M

How to use

ASR (child-voice Whisper)

from faster_whisper import WhisperModel
model = WhisperModel("DrUkachi/ktt-math-tutor-models",
                     device="cpu", compute_type="int8",
                     local_files_only=False)
segments, _ = model.transcribe(wav, language="en", beam_size=1)

Or, via the tutor's wrapper (auto-picks tutor/asr_model/ from the repo):

git clone https://github.com/DrUkachi/ktt-math-tutor
cd ktt-math-tutor && pip install -r requirements.txt
python demo.py

Eval on the in-distribution child-voice corpus (36 clips, pitched +3/+4.5/+6 semitones):

• Baseline vanilla Whisper-tiny int8: WER 0.7048
• This LoRA-tuned model: WER 0.0000

See scripts/eval_wer.py and metrics/wer_*.json in the code repo.

LLM head (weekly parent summary)

from llama_cpp import Llama
llm = Llama(
    model_path="tinyllama-numeracy-Q4_K_M.gguf",
    n_ctx=512, n_threads=4, verbose=False,
)
r = llm.create_chat_completion(messages=[
    {"role": "system", "content": "You are a warm math tutor. One short sentence."},
    {"role": "user", "content": "The child is strong at addition; needs practice on number sense."},
])

Or via the tutor's wrapper (tutor/llm_head.py): the model is resolved in order $TUTOR_LLM_GGUF → this tuned Q4_K_M → community TinyLlama base → deterministic fallback. None of the LLM path is in the inference hot path; it runs once per learner per week for the voiced parent summary.

Training recipes

• ASR LoRA: scripts/train_whisper_lora.py — 4 epochs on L4 GPU, LoRA r=16 on q_proj/v_proj, merge, export to CT2 int8.
• LLM QLoRA: scripts/train_llm_qlora.py — 2 epochs on L4 GPU, NF4 4-bit base, LoRA r=16 on q/k/v/o_proj, merge, convert to GGUF via pinned llama.cpp b4400 script, quantise to Q4_K_M via the llama_cpp.llama_model_quantize Python binding.

License

MIT. Attribution welcomed; not required.