README.md

metadata

license: gemma
datasets:
  - sam-paech/gutenberg3-generalfiction-scifi-fantasy-romance-adventure-dpo
  - sam-paech/gutenbergs_1_2_3_antislop-dpo
language:
  - en
base_model:
  - google/gemma-2-9b-it
library_name: transformers
tags:
  - creative-writing

Oblivion's End

A merged LoRA for gemma-2-9b-it, trained using DPO datasets for creative writing using my DPO training notebook.

Model Details

How to Use

from unsloth import FastLanguageModel # we use unsloth for faster inference
import torch
max_seq_length = 4096 
dtype = None
load_in_4bit = False

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = "mehmetkeremturkcan/oblivionsend",
    max_seq_length = max_seq_length,
    dtype = dtype,
    load_in_4bit = load_in_4bit
)

from transformers import TextStreamer
FastLanguageModel.for_inference(model)
text_streamer = TextStreamer(tokenizer)

inputs = tokenizer(
[
"""<start_of_turn>user
Write a story with the following description: Setting - a dark abandoned watchtower and its environs. A wizard carefully explores a tomb where a priest of a dark, dead God has raised a band of brigands that have been terrorizing a town."""+ """<end_of_turn>
<start_of_turn>model
"""
], return_tensors = "pt").to("cuda")

_ = model.generate(**inputs, streamer = text_streamer, max_new_tokens = 4096, num_beams=1, temperature=1.0, do_sample=True)

Model Description

• Finetuned from model: google/gemma-2-9b-it

Model Sources [optional]

• Repository: GitHub.

Uses

Made for creative writing.

Training Details

Training Data

Check out the model card details.

Training Procedure

Model training performance (margins) are available in the wandb instance.

Training Hyperparameters

• Training regime: bf16 on a 1x 80GB A100 node.

Environmental Impact

Total emissions are estimated to be 0.83 kgCO$_2$eq.