dolphin-v2-8b-abliterated/README.md

---
license: apache-2.0
base_model: Qwen/Qwen3-8B
tags:
  - uncensored
  - abliterated
  - qwen3
  - dolphin
  - sft
  - trc
language:
  - en
pipeline_tag: text-generation
model-index:
  - name: dolphin-v2-8b-abliterated
    results:
      - task:
          type: multiple-choice
          name: ARC Challenge
        dataset:
          name: ARC Challenge
          type: ai2_arc
          config: ARC-Challenge
          split: test
        metrics:
          - type: acc
            value: 56.5
            name: Accuracy
          - type: acc_norm
            value: 54.0
            name: Normalized Accuracy
      - task:
          type: multiple-choice
          name: HellaSwag
        dataset:
          name: HellaSwag
          type: Rowan/hellaswag
          split: validation
        metrics:
          - type: acc_norm
            value: 64.5
            name: Normalized Accuracy
      - task:
          type: multiple-choice
          name: TruthfulQA MC2
        dataset:
          name: TruthfulQA
          type: truthful_qa
          config: multiple_choice
          split: validation
        metrics:
          - type: acc
            value: 48.8
            name: Accuracy
      - task:
          type: multiple-choice
          name: Winogrande
        dataset:
          name: Winogrande
          type: winogrande
          config: winogrande_xl
          split: validation
        metrics:
          - type: acc
            value: 57.0
            name: Accuracy
---

# Dolphin V2 8B Abliterated

An uncensored 8B parameter language model built on Qwen3-8B, fine-tuned on 1.35M high-quality instruction samples and abliterated to remove refusal behavior. Developed for TRC (TPU Research Cloud) research.

## Model Details

- **Architecture:** Qwen3ForCausalLM (36 layers, 4096 hidden, 32 attn heads, 8 KV heads)
- **Parameters:** 8.2B
- **Context Length:** 4096 (trained), 40960 (max supported)
- **Precision:** bfloat16
- **License:** Apache 2.0

## Training

### SFT Phase
- **Base model:** Qwen/Qwen3-8B
- **Hardware:** Google Cloud TPU v6e-16 (spot)
- **Framework:** MaxText (JAX)
- **Steps:** 130,000 (~3 epochs)
- **Learning rate:** 5e-6 with cosine decay
- **Warmup:** 200 steps
- **Effective batch size:** 16
- **Sequence length:** 4096

### Training Dataset (1.35M samples)

| Dataset | Samples | Purpose |
|---------|---------|---------|
| NousResearch/Hermes-3-Dataset | ~959K | Core uncensored assistant behavior |
| allenai/tulu-3-sft-mixture | ~200K | Diverse instruction following |
| HuggingFaceTB/smoltalk (magpie-ultra) | ~100K | High quality diverse tasks |
| HuggingFaceTB/smoltalk (numina-cot) | ~50K | Math reasoning |
| HuggingFaceTB/smoltalk (self-oss-instruct) | ~50K | Code generation |
| LDJnr/Capybara | ~16K | Multi-turn conversations |

All data was filtered to remove refusal patterns, safety-alignment subsets, and `<think>` reasoning tags.

### Abliteration Phase

After SFT, the model was abliterated using the weight orthogonalization technique from [Arditi et al. (2024)](https://arxiv.org/abs/2406.11717) to remove residual refusal behavior.

- **Technique:** Multi-direction abliteration (weight orthogonalization)
- **Directions removed:** 5
- **Target layers:** 35, 34, 36, 33, 16 (selected by highest refusal direction scores)
- **Samples used:** 256 harmful/harmless instruction pairs
- **Method:** For each selected layer, the refusal direction was identified via mean difference between harmful and harmless activations, then orthogonalized out of the weight matrices.

## Benchmark Results

Evaluated using [lm-evaluation-harness](https://github.com/EleutherAI/lm-evaluation-harness) with 200 samples per task, 5-shot (except TruthfulQA which is 0-shot).

| Benchmark | Metric | Score |
|-----------|--------|-------|
| ARC-Challenge | acc | 56.5% |
| ARC-Challenge | acc_norm | 54.0% |
| HellaSwag | acc_norm | 64.5% |
| TruthfulQA MC2 | acc | 48.8% |
| Winogrande | acc | 57.0% |

## GGUF Quantizations

| File | Quant | Size | Description |
|------|-------|------|-------------|
| `dolphin-v2-8b-abliterated-Q8_0.gguf` | Q8_0 | 8.3 GB | Best quality quantization |
| `dolphin-v2-8b-abliterated-Q4_K_M.gguf` | Q4_K_M | 4.8 GB | Good balance of quality and size |

### Usage with llama.cpp

```bash
llama-server -m dolphin-v2-8b-abliterated-Q8_0.gguf -ngl 99 -c 4096
```

### Usage with Ollama

```bash
# Create a Modelfile
echo 'FROM ./dolphin-v2-8b-abliterated-Q8_0.gguf' > Modelfile
ollama create dolphin-v2-abliterated -f Modelfile
ollama run dolphin-v2-abliterated
```

### Usage with Transformers

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("0arch-io/dolphin-v2-8b-abliterated", torch_dtype="bfloat16", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained("0arch-io/dolphin-v2-8b-abliterated")

messages = [{"role": "user", "content": "Hello, how are you?"}]
inputs = tokenizer.apply_chat_template(messages, return_tensors="pt", add_generation_prompt=True).to(model.device)
outputs = model.generate(inputs, max_new_tokens=512, temperature=0.7)
print(tokenizer.decode(outputs[0][inputs.shape[-1]:], skip_special_tokens=True))
```

## Disclaimer

This is a research model with no content filters. It will comply with any request without refusing. The creators are not responsible for how this model is used. Use responsibly.

## Acknowledgments

- **Qwen team** for the Qwen3-8B base model
- **Google TRC** for TPU compute
- **NousResearch** for the Hermes-3 dataset
- **Arditi et al.** for the abliteration technique
- Built with [MaxText](https://github.com/AI-Hypercomputer/maxtext) on Google Cloud TPU