base_model: openbmb/MiniCPM5-2B
library_name: peft
license: apache-2.0
pipeline_tag: text-generation
tags:
- lora
- transformers
- abliterated
- fine-tuning
- uncensored
LoRA-Fine-Tuned Abliterated MiniCPM5-2B
A LoRA adapter that fine-tunes an abliterated (refusal-removed) MiniCPM5-2B model to improve response quality while maintaining refusal-free behavior.
Model Details
- Base model:
openbmb/MiniCPM5-2B(abliterated) - Adapter type: LoRA (r=16, alpha=32, dropout=0.1)
- Trainable params: 25.1M (1% of 2.5B total)
- Training steps: 2000
- Dataset: 5000 high-quality non-refusal training samples
- Precision: FP16
- GPU: AMD RX 9060 XT (ROCm 7.2)
Training
- Dataset: Curated non-refusal conversational + coding data (5000 samples)
- LR: 3e-4 with cosine annealing, 200 warmup steps
- Gradient accumulation: 8 steps (effective batch size 16)
- Training time: ~85 minutes on RX 9060 XT
- Final loss: ~0.05 (excellent convergence)
Results
The fine-tuned adapter significantly improves response quality:
| Task | Before (base ablated) | After (LoRA) |
|---|---|---|
| Factual QA | Didn't answer | ✅ "Paris" |
| Coding | Didn't answer | ✅ Working code |
| Math | Didn't answer | ✅ "2x" |
| Refusal check | Over-thought | ✅ Non-refusal |
Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
base = AutoModelForCausalLM.from_pretrained("openbmb/MiniCPM5-2B", torch_dtype="auto")
model = PeftModel.from_pretrained(base, "KidIkaros/abliterated-minicpm5-2b-lora-finetuned")
Notes
This is a LoRA adapter — it requires the abliterated base model to function. The base model weights are not included in this repo.
Contact
- GitHub: @KidIkaros