Fine-tuned Qwen2.5-0.5B-Instruct for routing student speech during live assignment defense oral assessments. Runs on CPU via llama-cpp-python — no GPU required.

Intents

Intent	Platform action
`answer`	Do nothing — student is answering (default)
`navigate_back`	Jump to `target_question_id` (e.g. `q2`)
`continue`	Resume forward after failed recall

Model details


Base model	Qwen/Qwen2.5-0.5B-Instruct
Fine-tuning	LoRA SFT (rank 8, ~4.4M trainable params)
Quantization	Q3_K_M
File size	~339 MB
RAM at inference	~410–470 MB (`n_ctx=1024`)
Eval token accuracy	99.2%

Files in this repository

File	Description
`voice-command-intent-qwen0.5b-Q3_K_M.gguf`	Production weights (ship this)
`inference/`	Python wrapper (`VoiceCommandClassifier`, prompts, schema, guard)

Quick start

Download GGUF only

python

from huggingface_hub import hf_hub_download

gguf = hf_hub_download(
    repo_id="yiyuliu/voice-command-intent-qwen0.5b",
    filename="voice-command-intent-qwen0.5b-Q3_K_M.gguf",
)

Full integration (GGUF + inference code)

Clone this repo or copy voice-command-intent-qwen0.5b-Q3_K_M.gguf + inference/ into your project.
Install dependencies:

bash

cd inference
pip install -e .

Classify an utterance:

python

from tinylm.inference import VoiceCommandClassifier
from tinylm.prompts import QuestionRecord

clf = VoiceCommandClassifier("voice-command-intent-qwen0.5b-Q3_K_M.gguf")

result = clf.classify(
    questions=[
        QuestionRecord("q1", "remember", "answered"),
        QuestionRecord("q2", "understand", "answered", focus="technical debt"),
        QuestionRecord("q3", "apply", "current"),
    ],
    current_question_id="q3",
    utterance="could I correct my previous question",
    subject="Computer Science",
)

print(result.to_action())
# {"action": "navigate", "target_question_id": "q2"}

Verify install

bash

cd inference
python smoke_test.py

Runtime inputs

Required: student utterance (STT text), Bloom session position (q1–q6, status, optional focus term).

Not required: full lecturer question text or assignment content.

Questions follow Bloom's taxonomy: remember → understand → apply → analyze → evaluate → create.

Environment variables (optional)

Variable	Default
`TINYLM_GGUF_PATH`	auto-resolve GGUF in repo root
`TINYLM_N_CTX`	`1024`
`TINYLM_MAX_TOKENS`	`64`

License

Fine-tuned weights: Apache 2.0 (same as base Qwen2.5)
Inference code: MIT

Base model license: Qwen/Qwen2.5-0.5B-Instruct