README.md

---
base_model: unsloth/GLM-4.7-Flash
language:
- en
license: other
tags:
- typo3
- finetuned
- lora
- peft
- unsloth
- glm
- cms
- php
---

# GLM-4.7-Flash — TYPO3 v13.4 / v14 Finetuned

A LoRA adapter finetuned on top of [GLM-4.7-Flash](https://huggingface.co/unsloth/GLM-4.7-Flash)
for TYPO3 CMS development — covering v13.4 LTS and v14.

## What it knows

Trained on **17,649 Q&A pairs** generated from the official TYPO3 documentation:

- **Core API** (TYPO3 Explained) — PSR-15 middleware, events, DI, site sets
- **TCA Reference** — all field types, column configs, type definitions
- **TypoScript Reference** — content objects, stdWrap, conditions, functions
- **Fluid Templates** — ViewHelpers, partials, layouts, standalone rendering
- **TSconfig Reference** — page and user TSconfig
- **Sitepackage Tutorial** — site package structure, configuration
- **Getting Started Tutorial** — installation, setup, basic concepts
- All sections covering both **v13.4 LTS** and **v14 (main)**

## Training details

| Parameter | Value |
|---|---|
| Base model | GLM-4.7-Flash (30B MoE, 3B active) |
| Method | LoRA (16-bit) |
| LoRA rank | 32 |
| LoRA alpha | 32 |
| Training steps | 900 |
| Batch size | 8 (1 × 8 grad accum) |
| Learning rate | 2e-4 (cosine schedule) |
| Final loss | 1.032 |
| Training time | ~113 minutes on A100 80GB |

## Usage

```python
from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    "renalpha/glm47-flash-typo3",
    max_seq_length = 2048,
    dtype = None,
    load_in_4bit = False,
)

messages = [
    {
        "role": "system",
        "content": "You are an expert TYPO3 13.4 developer."
    },
    {
        "role": "user", 
        "content": "How do I create a custom TCA field type in TYPO3 v13.4?"
    }
]

inputs = tokenizer.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    return_tensors="pt",
).to("cuda")

outputs = model.generate(
    input_ids=inputs,
    max_new_tokens=512,
    temperature=0.7,
    do_sample=True,
)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
```

## Dataset

Generated from the official TYPO3 GitHub documentation repos using
[GLM-4-32B-0414](https://huggingface.co/THUDM/glm-4-32b-0414) as the
synthetic data generator.

Repos used:
- `TYPO3-Documentation/TYPO3CMS-Reference-CoreApi`
- `TYPO3-Documentation/TYPO3CMS-Reference-TCA`
- `TYPO3-Documentation/TYPO3CMS-Reference-Typoscript`
- `TYPO3-Documentation/TYPO3CMS-Reference-ViewHelper`
- `TYPO3-Documentation/TYPO3CMS-Reference-TSconfig`
- `TYPO3-Documentation/TYPO3CMS-Tutorial-GettingStarted`
- `TYPO3-Documentation/TYPO3CMS-Tutorial-SitePackage`