Commit
•
096da7f
1
Parent(s):
f17b8c2
Update README.md
Browse files
README.md
CHANGED
@@ -4,4 +4,85 @@ tags:
|
|
4 |
- mistral
|
5 |
- roleplay
|
6 |
---
|
7 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
4 |
- mistral
|
5 |
- roleplay
|
6 |
---
|
7 |
+
|
8 |
+
Personal testing GGUF-IQ-Imatrix quants for [flammenai/flammen18X-mistral-7B](https://huggingface.co/flammenai/flammen18X-mistral-7B).
|
9 |
+
|
10 |
+
# Original model card information:
|
11 |
+
|
12 |
+
![image/png](https://huggingface.co/nbeerbower/flammen13X-mistral-7B/resolve/main/flammen13x.png)
|
13 |
+
|
14 |
+
# flammen18X-mistral-7B
|
15 |
+
|
16 |
+
A Mistral 7B LLM built from merging pretrained models and finetuning on [ResplendentAI/NSFW_RP_Format_DPO](https://huggingface.co/datasets/ResplendentAI/NSFW_RP_Format_DPO).
|
17 |
+
Flammen specializes in exceptional character roleplay, creative writing, and general intelligence
|
18 |
+
|
19 |
+
### Method
|
20 |
+
|
21 |
+
Finetuned using an A100 on Google Colab.
|
22 |
+
|
23 |
+
[Fine-tune a Mistral-7b model with Direct Preference Optimization](https://towardsdatascience.com/fine-tune-a-mistral-7b-model-with-direct-preference-optimization-708042745aac) - [Maxime Labonne](https://huggingface.co/mlabonne)
|
24 |
+
|
25 |
+
### Configuration
|
26 |
+
|
27 |
+
LoRA, model, and training settings:
|
28 |
+
|
29 |
+
```python
|
30 |
+
# LoRA configuration
|
31 |
+
peft_config = LoraConfig(
|
32 |
+
r=16,
|
33 |
+
lora_alpha=16,
|
34 |
+
lora_dropout=0.05,
|
35 |
+
bias="none",
|
36 |
+
task_type="CAUSAL_LM",
|
37 |
+
target_modules=['k_proj', 'gate_proj', 'v_proj', 'up_proj', 'q_proj', 'o_proj', 'down_proj']
|
38 |
+
)
|
39 |
+
|
40 |
+
# Model to fine-tune
|
41 |
+
model = AutoModelForCausalLM.from_pretrained(
|
42 |
+
model_name,
|
43 |
+
torch_dtype=torch.bfloat16,
|
44 |
+
load_in_4bit=True
|
45 |
+
)
|
46 |
+
model.config.use_cache = False
|
47 |
+
|
48 |
+
# Reference model
|
49 |
+
ref_model = AutoModelForCausalLM.from_pretrained(
|
50 |
+
model_name,
|
51 |
+
torch_dtype=torch.bfloat16,
|
52 |
+
load_in_4bit=True
|
53 |
+
)
|
54 |
+
|
55 |
+
# Training arguments
|
56 |
+
training_args = TrainingArguments(
|
57 |
+
per_device_train_batch_size=2,
|
58 |
+
gradient_accumulation_steps=8,
|
59 |
+
gradient_checkpointing=True,
|
60 |
+
learning_rate=5e-5,
|
61 |
+
lr_scheduler_type="cosine",
|
62 |
+
max_steps=420,
|
63 |
+
save_strategy="no",
|
64 |
+
logging_steps=1,
|
65 |
+
output_dir=new_model,
|
66 |
+
optim="paged_adamw_32bit",
|
67 |
+
warmup_steps=100,
|
68 |
+
bf16=True,
|
69 |
+
report_to="wandb",
|
70 |
+
)
|
71 |
+
|
72 |
+
# Create DPO trainer
|
73 |
+
dpo_trainer = DPOTrainer(
|
74 |
+
model,
|
75 |
+
ref_model,
|
76 |
+
args=training_args,
|
77 |
+
train_dataset=dataset,
|
78 |
+
tokenizer=tokenizer,
|
79 |
+
peft_config=peft_config,
|
80 |
+
beta=0.1,
|
81 |
+
max_prompt_length=1024,
|
82 |
+
max_length=1536,
|
83 |
+
force_use_ref_model=True
|
84 |
+
)
|
85 |
+
|
86 |
+
# Fine-tune model with DPO
|
87 |
+
dpo_trainer.train()
|
88 |
+
```
|