111 lines
3.4 KiB
Markdown
111 lines
3.4 KiB
Markdown
---
|
|
license: apache-2.0
|
|
language:
|
|
- en
|
|
pipeline_tag: text-generation
|
|
dtype: bfloat16
|
|
tags:
|
|
- merge
|
|
---
|
|
|
|
# EDIT:
|
|
Always check my space for the latest benchmark results for my models!
|
|
* https://huggingface.co/spaces/CultriX/Yet_Another_LLM_Leaderboard
|
|
|
|
# Results:
|
|
T: 🟦
|
|
Model: CultriX/MistralTrix-v1 📑
|
|
Average: 73.39
|
|
ARC: 72.27
|
|
HellaSwag: 88.33
|
|
MMLU: 65.24
|
|
TruthfulQA: 70.73
|
|
Winogrande: 80.98
|
|
GSM8K: 62.77
|
|
|
|
# Edit/Disclaimer:
|
|
Currently the #1 ranked 7B LLM on the LLM Leaderboards, woah!
|
|
I did not expect that result at all and am in no way a professional when it comes to LLM's or computer science in general,
|
|
just a guy that likes to nerd about and tinker around.
|
|
|
|
For those wondering how I achieved this, the answer is that I simply attempted to apply the techniques outlined in this amazing article myself: https://towardsdatascience.com/fine-tune-a-mistral-7b-model-with-direct-preference-optimization-708042745aac
|
|
Therefore, all credit basically goes to the guy who wrote that.
|
|
He offers the exact Colab notebook I used to train this model for free, as well as a really nice GitHub page I hope he doesn't mind me sharing: https://github.com/mlabonne/llm-course/
|
|
So huge thank you to him for sharing his knowledge and learning me a thing or two in the process!
|
|
|
|
# GGUF
|
|
I attempted to quantisize the model myself, which again I pretty much have no clue about, but it seems to run fine for me when I test them:
|
|
https://huggingface.co/CultriX/MistralTrix-v1-GGUF
|
|
|
|
I'll say it one more time though:
|
|
"I am a complete beginner to all of this, so if these do end up sucking don't be surprised."
|
|
|
|
You have been warned :)
|
|
|
|
# Description:
|
|
(trained on a single Colab GPU in less than a few hours)
|
|
|
|
MistralTrix-v1 is an zyh3826/GML-Mistral-merged-v1 model that has been further fine-tuned with Direct Preference Optimization (DPO) using Intel's dataset for neural-chat-7b-v3-1.
|
|
It surpasses the original model on several benchmarks (see results).
|
|
|
|
It is directly inspired by the RLHF process described by Intel/neural-chat-7b-v3-1's authors to improve performance.
|
|
I used the same dataset and reformatted it to apply the ChatML template.
|
|
|
|
The code to train this model is available on Google Colab and GitHub.
|
|
Fine-tuning took about an hour on Google Colab A-1000 GPU with 40GB VRAM.
|
|
|
|
# TRAINING SPECIFICATIONS
|
|
> 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.float16,
|
|
load_in_4bit=True
|
|
)
|
|
model.config.use_cache = False
|
|
|
|
> Reference model
|
|
ref_model = AutoModelForCausalLM.from_pretrained(
|
|
model_name,
|
|
torch_dtype=torch.float16,
|
|
load_in_4bit=True
|
|
)
|
|
|
|
> Training arguments
|
|
training_args = TrainingArguments(
|
|
per_device_train_batch_size=4,
|
|
gradient_accumulation_steps=4,
|
|
gradient_checkpointing=True,
|
|
learning_rate=5e-5,
|
|
lr_scheduler_type="cosine",
|
|
max_steps=200,
|
|
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=1024,
|
|
max_length=1536,
|
|
) |