Downloads · 30 days
145
5% of all-time downloads
RichardErkhov/flammenai_-_flammen23X-mistral-7B-gguf
flammenai_-_flammen23X-mistral-7B-gguf is a machine learning model from RichardErkhov. Use it for the machine learning task on the model card, and read the license before you ship it in a product.
Downloads · 30 days
145
5% of all-time downloads
All-time downloads
3.1K
Public
Repo size
62 GB
Likes
0
Public
Click a slice to open those files.
.gguf65.5 GB · 100%
From the Hugging Face model README
Quantization made by Richard Erkhov.
flammen23X-mistral-7B - GGUF
| Name | Quant method | Size |
|---|---|---|
| flammen23X-mistral-7B.Q2_K.gguf | Q2_K | 2.53GB |
| flammen23X-mistral-7B.IQ3_XS.gguf | IQ3_XS | 2.81GB |
| flammen23X-mistral-7B.IQ3_S.gguf | IQ3_S | 2.96GB |
| flammen23X-mistral-7B.Q3_K_S.gguf | Q3_K_S | 2.95GB |
| flammen23X-mistral-7B.IQ3_M.gguf | IQ3_M | 3.06GB |
| flammen23X-mistral-7B.Q3_K.gguf | Q3_K | 3.28GB |
| flammen23X-mistral-7B.Q3_K_M.gguf | Q3_K_M | 3.28GB |
| flammen23X-mistral-7B.Q3_K_L.gguf | Q3_K_L | 3.56GB |
| flammen23X-mistral-7B.IQ4_XS.gguf | IQ4_XS | 3.67GB |
| flammen23X-mistral-7B.Q4_0.gguf | Q4_0 | 3.83GB |
| flammen23X-mistral-7B.IQ4_NL.gguf | IQ4_NL | 3.87GB |
| flammen23X-mistral-7B.Q4_K_S.gguf | Q4_K_S | 3.86GB |
| flammen23X-mistral-7B.Q4_K.gguf | Q4_K | 2.87GB |
| flammen23X-mistral-7B.Q4_K_M.gguf | Q4_K_M | 2.09GB |
| flammen23X-mistral-7B.Q4_1.gguf | Q4_1 | 3.8GB |
| flammen23X-mistral-7B.Q5_0.gguf | Q5_0 | 2.08GB |
| flammen23X-mistral-7B.Q5_K_S.gguf | Q5_K_S | 1.0GB |
| flammen23X-mistral-7B.Q5_K.gguf | Q5_K | 0.43GB |
| flammen23X-mistral-7B.Q5_K_M.gguf | Q5_K_M | 4.78GB |
| flammen23X-mistral-7B.Q5_1.gguf | Q5_1 | 2.23GB |
| flammen23X-mistral-7B.Q6_K.gguf | Q6_K | 0.9GB |
| flammen23X-mistral-7B.Q8_0.gguf | Q8_0 | 1.18GB |
library_name: transformers license: apache-2.0 base_model:

A Mistral 7B LLM built from merging pretrained models and finetuning on flammenai/character-roleplay-DPO. Flammen specializes in exceptional character roleplay, creative writing, and general intelligence
Finetuned using an A100 on Google Colab.
Fine-tune a Mistral-7b model with Direct Preference Optimization - Maxime Labonne
System prompt, dataset formatting:
def chatml_format(example):
# Format system
#system = ""
systemMessage = "Write a character roleplay dialogue using asterisk roleplay format based on the following character descriptions and scenario. (Each line in your response must be from the perspective of one of these characters)"
system = "<|im_start|>system\n" + systemMessage + "<|im_end|>\n"
# Format instruction
prompt = "<|im_start|>user\n" + example['input'] + "<|im_end|>\n<|im_start|>assistant\n"
# Format chosen answer
chosen = example['output'] + "<|im_end|>\n"
# Format rejected answer
rejected = example['rejected'] + "<|im_end|>\n"
return {
"prompt": system + prompt,
"chosen": chosen,
"rejected": rejected,
}
dataset = load_dataset("flammenai/character-roleplay-DPO")['train']
# Save columns
original_columns = dataset.column_names
# Tokenizer
tokenizer = AutoTokenizer.from_pretrained(model_name)
tokenizer.pad_token = tokenizer.eos_token
tokenizer.padding_side = "left"
# Format dataset
dataset = dataset.map(
chatml_format,
remove_columns=original_columns
)
LoRA, model, and training settings:
# LoRA configuration
peft_config = LoraConfig(
r=16,
lora_alpha=16,
lora_dropout=0.05,
bias="none",
task_type="CAUSAL_LM",
target_modules=['k_proj', 'gate_proj', 'v_proj', 'up_proj', 'q_proj', 'o_proj', 'down_proj']
)
# Model to fine-tune
model = AutoModelForCausalLM.from_pretrained(
model_name,
torch_dtype=torch.bfloat16,
load_in_4bit=True
)
model.config.use_cache = False
# Reference model
ref_model = AutoModelForCausalLM.from_pretrained(
model_name,
torch_dtype=torch.bfloat16,
load_in_4bit=True
)
# Training arguments
training_args = TrainingArguments(
per_device_train_batch_size=2,
gradient_accumulation_steps=4,
gradient_checkpointing=True,
learning_rate=5e-5,
lr_scheduler_type="cosine",
max_steps=350,
save_strategy="no",
logging_steps=1,
output_dir=new_model,
optim="paged_adamw_32bit",
warmup_steps=100,
bf16=True,
report_to="wandb",
)
# Create DPO trainer
dpo_trainer = DPOTrainer(
model,
ref_model,
args=training_args,
train_dataset=dataset,
tokenizer=tokenizer,
peft_config=peft_config,
beta=0.1,
max_prompt_length=4096,
max_length=8192,
force_use_ref_model=True
)