README.md

---
base_model: google/gemma-4-e4b-it
library_name: peft
license: mit
tags:
  - lora
  - peft
  - gemma-4
  - education
  - lesson-planning
  - tlc
language:
  - en
---

# TLC — Christine LoRA (Gemma 4 E4B)

QLoRA adapter for **Gemma 4 E4B** trained to play the **Christine**
persona inside [TLC — Teacher's Lesson
Creator](https://github.com/hardcoded74/tlc), an open-source
lesson-building tool for the Kaggle Gemma 4 Good Hackathon
(Impact Track).

Christine is one of two collaborating Teacher's Assistants.
**Christine owns depth and engagement** — the engagement moment,
hands-on demonstrations, teacher delivery notes, discussion prompts,
differentiation, misconceptions. Hunter ([sibling
adapter](https://huggingface.co/hardcoded74/tlc-gemma-4-e4b-hunter-lora))
owns structure and rigor; 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` (engagement
hooks, demonstrations, vocabulary with examples, misconceptions with
corrections, 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 Christine's pedagogical voice.

## Quick start

```python
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-christine-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`](https://github.com/hardcoded74/tlc/blob/main/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/`](https://github.com/hardcoded74/tlc/tree/main/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 Hunter adapter
and hot-swapped per request. Using it solo will work but gives you
only the depth/engagement half of TLC's output. The deterministic
merge in [`lib/merge.ts`](https://github.com/hardcoded74/tlc/blob/main/lib/merge.ts)
combines both into a single lesson package.

## License

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

## Citation

```bibtex
@misc{tlc-christine-2026,
  title  = {TLC Christine — Gemma 4 E4B QLoRA for K-12 lesson depth and engagement},
  author = {Sam},
  year   = {2026},
  url    = {https://huggingface.co/hardcoded74/tlc-gemma-4-e4b-christine-lora}
}
```