README.md

---
library_name: transformers
tags:
- mergekit
- merge
- llama-3.1
- roleplay
- function calling
base_model:
- T145/ZEUS-8B-V2
license: llama3.1
model-index:
- name: ZEUS-8B-V2-abliterated
  results:
  - task:
      type: text-generation
      name: Text Generation
    dataset:
      name: IFEval (0-Shot)
      type: wis-k/instruction-following-eval
      split: train
      args:
        num_few_shot: 0
    metrics:
    - type: inst_level_strict_acc and prompt_level_strict_acc
      value: 78.95
      name: averaged accuracy
    source:
      url: https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard#/?search=T145%2FZEUS-8B-V2-abliterated
      name: Open LLM Leaderboard
  - task:
      type: text-generation
      name: Text Generation
    dataset:
      name: BBH (3-Shot)
      type: SaylorTwift/bbh
      split: test
      args:
        num_few_shot: 3
    metrics:
    - type: acc_norm
      value: 30.98
      name: normalized accuracy
    source:
      url: https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard#/?search=T145%2FZEUS-8B-V2-abliterated
      name: Open LLM Leaderboard
  - task:
      type: text-generation
      name: Text Generation
    dataset:
      name: MATH Lvl 5 (4-Shot)
      type: lighteval/MATH-Hard
      split: test
      args:
        num_few_shot: 4
    metrics:
    - type: exact_match
      value: 20.62
      name: exact match
    source:
      url: https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard#/?search=T145%2FZEUS-8B-V2-abliterated
      name: Open LLM Leaderboard
  - task:
      type: text-generation
      name: Text Generation
    dataset:
      name: GPQA (0-shot)
      type: Idavidrein/gpqa
      split: train
      args:
        num_few_shot: 0
    metrics:
    - type: acc_norm
      value: 8.39
      name: acc_norm
    source:
      url: https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard#/?search=T145%2FZEUS-8B-V2-abliterated
      name: Open LLM Leaderboard
  - task:
      type: text-generation
      name: Text Generation
    dataset:
      name: MuSR (0-shot)
      type: TAUR-Lab/MuSR
      args:
        num_few_shot: 0
    metrics:
    - type: acc_norm
      value: 7.92
      name: acc_norm
    source:
      url: https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard#/?search=T145%2FZEUS-8B-V2-abliterated
      name: Open LLM Leaderboard
  - task:
      type: text-generation
      name: Text Generation
    dataset:
      name: MMLU-PRO (5-shot)
      type: TIGER-Lab/MMLU-Pro
      config: main
      split: test
      args:
        num_few_shot: 5
    metrics:
    - type: acc
      value: 31.39
      name: accuracy
    source:
      url: https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard#/?search=T145%2FZEUS-8B-V2-abliterated
      name: Open LLM Leaderboard
---

# ZEUS 8B 🌩️ V2 - ABLITERATED

V2 abliterated using the following script:

```python
import gc
import random

import torch
from tqdm import tqdm
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig

MODEL_ID = "T145/ZEUS-8B-V2"

# More samples can help find the direction better.
NUM_PROMPT_SAMPLES = 32

# Used to skip the first and last layers for the modifications.
SKIP_BEGIN_LAYERS = 1
SKIP_END_LAYERS = 1

# The layer we will use for the refusal_dir calculation will be floor(LAYER_FRACTION_TO_USE * model.layers).
LAYER_FRACTION_TO_USE = 0.6

# Use a negative scale_factor to "induce" and a positive scale_factor of < 1 to "ablate" less.
SCALE_FACTOR = 1.0

torch.inference_mode()
torch.set_default_device("cpu")
torch.set_grad_enabled(False)

# Load the model on the GPU in quantized type if we can.
model = AutoModelForCausalLM.from_pretrained(
    MODEL_ID,
    trust_remote_code=True,
    torch_dtype=torch.float16,
    quantization_config=BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_compute_dtype=torch.float16),
    low_cpu_mem_usage=True,
    device_map='auto'
)
model.requires_grad_(False)

tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, trust_remote_code=True)
layer_idx = int(len(model.model.layers) * LAYER_FRACTION_TO_USE)

print("Layer index for refusal direction: " + str(layer_idx))

with open("harmful.txt", "r", encoding="utf-8") as f:
    harmful = f.readlines()

with open("harmless.txt", "r", encoding="utf-8") as f:
    harmless = f.readlines()

harmful_instructions = random.sample(harmful, min(NUM_PROMPT_SAMPLES, len(harmful)))
harmless_instructions = random.sample(harmless, min(NUM_PROMPT_SAMPLES, len(harmless)))

harmful_toks = [
    tokenizer.apply_chat_template(conversation=[{"role": "user", "content": insn}], add_generation_prompt=True, tokenize=False,
                                  return_tensors="pt") for insn in harmful_instructions]
harmless_toks = [
    tokenizer.apply_chat_template(conversation=[{"role": "user", "content": insn}], add_generation_prompt=True, tokenize=False,
                                  return_tensors="pt") for insn in harmless_instructions]

bar_generate = tqdm(total = len(harmful_instructions) + len(harmless_instructions), desc = "Generating samples")

# Only return the final hidden state of the layer we care about, and use 'cpu' to save VRAM.
def generate(toks):
    inputs = tokenizer(toks, return_tensors="pt", padding=True)
    inputs = inputs.to(model.device)
    output = model.generate(
        inputs['input_ids'],
        use_cache=False,
        max_new_tokens=1,
        return_dict_in_generate=True,
        output_hidden_states=True,
        attention_mask=inputs["attention_mask"],
        pad_token_id=tokenizer.eos_token_id
    )
    bar_generate.update(n=1)
    return output.hidden_states[0][layer_idx][:, -1, :].to('cpu') # Final hidden state = -1.

harmful_hidden = [generate(toks) for toks in harmful_toks]
harmless_hidden = [generate(toks) for toks in harmless_toks]

bar_generate.close()

harmful_mean = torch.stack(harmful_hidden).mean(dim=0)
harmless_mean = torch.stack(harmless_hidden).mean(dim=0)

refusal_dir = harmful_mean - harmless_mean
refusal_dir = refusal_dir.squeeze() / refusal_dir.norm()

torch.save(refusal_dir, MODEL_ID.replace("/", "_") + "_refusal_dir.pt")

# Free memory
del model
gc.collect()
torch.cuda.empty_cache()

# Reload the model in CPU memory with bfloat16 data type
model = AutoModelForCausalLM.from_pretrained(
    MODEL_ID,
    trust_remote_code=True,
    torch_dtype=torch.bfloat16,
    low_cpu_mem_usage=True,
    device_map='cpu'
)
model.requires_grad_(False)

# Make sure it's on the 'cpu' device.
if refusal_dir.device != model.device:
    refusal_dir = refusal_dir.to(model.device)

# Get the language model component and check it's as expected.
lm_model = model.model
assert hasattr(lm_model, 'layers'), "The model does not have the expected structure."

# Check the ranges are valid.
num_layers = len(lm_model.layers)
assert SKIP_BEGIN_LAYERS >= 0, "SKIP_BEGIN_LAYERS must be >= 0."
assert SKIP_END_LAYERS >= 0, "SKIP_END_LAYERS must be >= 0."
assert SKIP_BEGIN_LAYERS + SKIP_END_LAYERS < num_layers, "SKIP_BEGIN_LAYERS + SKIP_END_LAYERS must be < num_layers."

bar_layers = tqdm(total= (num_layers - (SKIP_BEGIN_LAYERS + SKIP_END_LAYERS)) * 2, desc = "Modifying tensors")

# NOTE: Use a negative scale_factor to "induce" and a positive scale_factor of < 1 to "ablate" less.
def modify_tensor(tensor_data, refusal_dir, scale_factor: float = 1.0):
    assert scale_factor <= 1.0, "Using a scale_factor of > 1 doesn't make sense..."
    tensor_float = tensor_data.to(torch.bfloat16)
    refusal_dir_float = refusal_dir.to(torch.bfloat16)
    tensor_float -= scale_factor * torch.matmul(torch.outer(refusal_dir_float, refusal_dir_float), tensor_float)
    tensor_modified = tensor_float.to(torch.bfloat16)
    bar_layers.update(1)
    return torch.nn.Parameter(tensor_modified)

# Modify the 'self_attn.o_proj.weight' and 'mlp.down_proj.weight' in each chosen layer.
# NOTE: These tensors names are speific to "llama" and may need changing.
#       - See here for others: https://github.com/arcee-ai/mergekit/tree/main/mergekit/_data/architectures
for layer_idx in range(SKIP_BEGIN_LAYERS, num_layers - SKIP_END_LAYERS):
    lm_model.layers[layer_idx].self_attn.o_proj.weight = modify_tensor(
        lm_model.layers[layer_idx].self_attn.o_proj.weight.data, refusal_dir, SCALE_FACTOR
    )
    lm_model.layers[layer_idx].mlp.down_proj.weight = modify_tensor(
        lm_model.layers[layer_idx].mlp.down_proj.weight.data, refusal_dir, SCALE_FACTOR
    )

bar_layers.close()

print("Saving modified model (with original tokenizer)...")

FIXED_ID = f"{MODEL_ID}-abliterated"
model.save_pretrained(FIXED_ID)
tokenizer.save_pretrained(FIXED_ID)
```

According to the script, **layer 19** is the primary target for abliteration.

# [Open LLM Leaderboard Evaluation Results](https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard)
Detailed results can be found [here](https://huggingface.co/datasets/open-llm-leaderboard/T145__ZEUS-8B-V2-abliterated-details)!
Summarized results can be found [here](https://huggingface.co/datasets/open-llm-leaderboard/contents/viewer/default/train?q=T145%2FZEUS-8B-V2-abliterated&sort[column]=Average%20%E2%AC%86%EF%B8%8F&sort[direction]=desc)!

|      Metric       |Value (%)|
|-------------------|--------:|
|**Average**        |    29.71|
|IFEval (0-Shot)    |    78.95|
|BBH (3-Shot)       |    30.98|
|MATH Lvl 5 (4-Shot)|    20.62|
|GPQA (0-shot)      |     8.39|
|MuSR (0-shot)      |     7.92|
|MMLU-PRO (5-shot)  |    31.39|