hardcoded74/tlc-gemma-4-e4b-hunter-lora

TLC — Hunter LoRA (Gemma 4 E4B)

QLoRA adapter for Gemma 4 E4B trained to play the Hunter persona inside TLC — Teacher's Lesson Creator, an open-source lesson-building tool for the Kaggle Gemma 4 Good Hackathon (Impact Track).

Hunter is one of two collaborating Teacher's Assistants. Hunter owns structure and rigor — the learning objective, the lesson sequence, the assessment, the answer key, the time math, standards alignment. Christine (sibling adapter) owns depth and engagement; the two outputs are merged by deterministic field-ownership rules at Phase 3.

What this adapter is for

TLC requires the model to emit strict JSON via Gemma 4's native function-calling. Stock Gemma 4 E4B can produce loose JSON but struggles to hit a deeply nested PersonaScaffoldSchema (lesson steps, materials, assessment, vocabulary, misconceptions, source provenance per field) from zero-shot. With ~250 schema-validated golden outputs synthesized from cloud Gemma 4 31B as teacher, this adapter learns the strict-schema discipline end-to-end while keeping Hunter's structural voice.

Quick start

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = AutoModelForCausalLM.from_pretrained("google/gemma-4-e4b-it")
tok = AutoTokenizer.from_pretrained("google/gemma-4-e4b-it")
model = PeftModel.from_pretrained(base, "hardcoded74/tlc-gemma-4-e4b-hunter-lora")

For llama.cpp serving (the way TLC actually deploys it), convert the adapter to GGUF via llama.cpp/convert_lora_to_gguf.py and load with --lora. See TLC's scripts/run_local_llama.sh for the exact serving config (Hunter at adapter id 0, Christine at id 1, both loaded with --lora-init-without-apply so the worker can hot-swap per request).

Training details

• Base: google/gemma-4-e4b-it
• Method: SFT (TRL) over QLoRA, NF4 base + bf16 compute
• Data: ~250 schema-validated PersonaScaffoldSchema outputs generated by Gemma 4 31B (dense) acting as the teacher model on a curated K-12 topic x grade matrix
• Hardware: Intel Arc B570 (10 GB) via Intel Extension for PyTorch
• Pipeline: Fully reproducible from training/ in the TLC repo — topic matrix, data-gen script, Arc training notebook, GGUF conversion

Pairing

This adapter is designed to be served alongside the Christine adapter and hot-swapped per request. Using it solo will work but gives you only the structural half of TLC's output. The deterministic merge in lib/merge.ts combines both into a single lesson package.

License

MIT. Fork it, build on it, improve it.

Citation

@misc{tlc-hunter-2026,
  title  = {TLC Hunter — Gemma 4 E4B QLoRA for K-12 lesson structure},
  author = {Sam},
  year   = {2026},
  url    = {https://huggingface.co/hardcoded74/tlc-gemma-4-e4b-hunter-lora}
}