|
|
--- |
|
|
license: apache-2.0 |
|
|
base_model: Qwen/Qwen3-4B-Instruct-2507 |
|
|
tags: |
|
|
- projectforty2 |
|
|
- tce-trained |
|
|
- alignment |
|
|
- coder |
|
|
- negative-alignment-tax |
|
|
--- |
|
|
|
|
|
# ford_prefect |
|
|
|
|
|
This model was trained using the **ProjectForty2 TCE** (Training & Calibration Environment). |
|
|
|
|
|
## Training Details |
|
|
|
|
|
- **Base Model**: Qwen/Qwen3-4B-Instruct-2507 |
|
|
- **Recipe**: coder |
|
|
- **Training Method**: LoRA fine-tuning with isotope-based alignment |
|
|
|
|
|
|
|
|
## What is TCE? |
|
|
|
|
|
The TCE is part of ProjectFort2, which provides tools for fine-tuning language models with specific behavioral "isotopes" - carefully crafted training examples that teach models epistemic humility, calibrated uncertainty, and other alignment properties. |
|
|
|
|
|
### Key Features: |
|
|
- **Negative Alignment Tax**: Training improves both safety AND capability metrics |
|
|
- **Isotope-based Training**: Modular behavioral components that can be combined |
|
|
- **Comprehensive Benchmarking**: TruthfulQA, MMLU, HumanEval, and more |
|
|
|
|
|
## Usage |
|
|
|
|
|
```python |
|
|
from transformers import AutoModelForCausalLM, AutoTokenizer |
|
|
from peft import PeftModel |
|
|
|
|
|
# Load base model |
|
|
base_model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-4B-Instruct-2507") |
|
|
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-4B-Instruct-2507") |
|
|
|
|
|
# Load LoRA adapter |
|
|
model = PeftModel.from_pretrained(base_model, "ProjectForty2/ford_prefect") |
|
|
``` |
|
|
|
|
|
## License |
|
|
|
|
|
Apache 2.0 |
|
|
|
|
|
## Links |
|
|
|
|
|
- [ProjectForty2](http://projectforty2.ai) |
