This is a LongWriter-llama3.1-8B fine-tune, produced through P-E-W's Heretic (v1.1.0) abliteration engine merged with the Magnitude-Preserving Orthogonal Ablation PR.

Note: This model was generated to provide data for redaihf in testing MPOA abliterations against standart abliterations.

Heretication Results

Score Metric	Value	Parameter	Value
Refusals	9/100	direction_index	per layer
KL Divergence	0.0743	attn.o_proj.max_weight	1.26
Initial Refusals	99/100	attn.o_proj.max_weight_position	20.09
		attn.o_proj.min_weight	1.09
		attn.o_proj.min_weight_distance	10.32
		mlp.down_proj.max_weight	1.48
		mlp.down_proj.max_weight_position	23.44
		mlp.down_proj.min_weight	1.25
		mlp.down_proj.min_weight_distance	15.65

Degree of Heretication

The Heresy Index weighs the resulting model's corruption by the process (KL Divergence) and its abolition of doctrine (Refusals) for a final verdict in classification.

Index Entry	Classification	Analysis
	Absolute Heresy	Less than 10/100 Refusals and 0.10 KL Divergence
	Tainted Heresy	Around 25-11/100 Refusals and/or -0.20-0.11 KL Divergence
	Impotent Heresy	Anything above 25/100 Refusals and 0.21 KL Divergence

Note: This is an arbitrary classification inspired by Warhammer 40K, having no tangible indication towards the model's performance.

LongWriter-llama3.1-8b

🤗 [LongWriter Dataset] • 💻 [Github Repo] • 📃 [LongWriter Paper]

LongWriter-llama3.1-8b is trained based on Meta-Llama-3.1-8B, and is capable of generating 10,000+ words at once.

Environment: transformers>=4.43.0

Please ahere to the prompt template (system prompt is optional): <<SYS>>\n{system prompt}\n<</SYS>>\n\n[INST]{query1}[/INST]{response1}[INST]{query2}[/INST]{response2}...

A simple demo for deployment of the model:

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch
tokenizer = AutoTokenizer.from_pretrained("THUDM/LongWriter-llama3.1-8b", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("THUDM/LongWriter-llama3.1-8b", torch_dtype=torch.bfloat16, trust_remote_code=True, device_map="auto")
model = model.eval()
query = "Write a 10000-word China travel guide"
prompt = f"[INST]{query}[/INST]"
input = tokenizer(prompt, truncation=False, return_tensors="pt").to(device)
context_length = input.input_ids.shape[-1]
output = model.generate(
    **input,
    max_new_tokens=32768,
    num_beams=1,
    do_sample=True,
    temperature=0.5,
)[0]
response = tokenizer.decode(output[context_length:], skip_special_tokens=True)
print(response)

You can also deploy the model with vllm, which allows 10,000+ words generation within a minute. Here is an example code:

model = LLM(
    model= "THUDM/LongWriter-llama3.1-8b",
    dtype="auto",
    trust_remote_code=True,
    tensor_parallel_size=1,
    max_model_len=32768,
    gpu_memory_utilization=0.5,
)
tokenizer = model.get_tokenizer()
generation_params = SamplingParams(
    temperature=0.5,
    top_p=0.8,
    top_k=50,
    max_tokens=32768,
    repetition_penalty=1,
)
query = "Write a 10000-word China travel guide"
prompt = f"[INST]{query}[/INST]"
input_ids = tokenizer(prompt, truncation=False, return_tensors="pt").input_ids[0].tolist()
outputs = model.generate(
    sampling_params=generation_params,
    prompt_token_ids=[input_ids],
)
output = outputs[0]
print(output.outputs[0].text)

License: Llama-3.1 License

Citation

If you find our work useful, please consider citing LongWriter:

@article{bai2024longwriter,
  title={LongWriter: Unleashing 10,000+ Word Generation from Long Context LLMs}, 
  author={Yushi Bai and Jiajie Zhang and Xin Lv and Linzhi Zheng and Siqi Zhu and Lei Hou and Yuxiao Dong and Jie Tang and Juanzi Li},
  journal={arXiv preprint arXiv:2408.07055},
  year={2024}
}