File size: 6,055 Bytes
0515aa9 86b1adc 22b829e 86b1adc 22b829e 6a13434 0515aa9 86b1adc cd2fa5c 86b1adc 2526e6e 86b1adc 22b829e 86b1adc 22b829e cd2fa5c 86b1adc 2526e6e 86b1adc 22b829e 0ef9766 86b1adc 22b829e 86b1adc 22b829e 0ef9766 6a13434 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 |
---
license: apache-2.0
tags:
- mlabonne/NeuralMarcoro14-7B
- dpo
- 7B
- winograd
- mistral
datasets:
- hromi/winograd_dpo_basic
base_model: mlabonne/NeuralMarcoro14-7B
model-index:
- name: Garrulus
results:
- task:
type: text-generation
name: Text Generation
dataset:
name: AI2 Reasoning Challenge (25-Shot)
type: ai2_arc
config: ARC-Challenge
split: test
args:
num_few_shot: 25
metrics:
- type: acc_norm
value: 73.29
name: normalized accuracy
source:
url: https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=udkai/Garrulus
name: Open LLM Leaderboard
- task:
type: text-generation
name: Text Generation
dataset:
name: HellaSwag (10-Shot)
type: hellaswag
split: validation
args:
num_few_shot: 10
metrics:
- type: acc_norm
value: 88.87
name: normalized accuracy
source:
url: https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=udkai/Garrulus
name: Open LLM Leaderboard
- task:
type: text-generation
name: Text Generation
dataset:
name: MMLU (5-Shot)
type: cais/mmlu
config: all
split: test
args:
num_few_shot: 5
metrics:
- type: acc
value: 64.57
name: accuracy
source:
url: https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=udkai/Garrulus
name: Open LLM Leaderboard
- task:
type: text-generation
name: Text Generation
dataset:
name: TruthfulQA (0-shot)
type: truthful_qa
config: multiple_choice
split: validation
args:
num_few_shot: 0
metrics:
- type: mc2
value: 68.23
source:
url: https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=udkai/Garrulus
name: Open LLM Leaderboard
- task:
type: text-generation
name: Text Generation
dataset:
name: Winogrande (5-shot)
type: winogrande
config: winogrande_xl
split: validation
args:
num_few_shot: 5
metrics:
- type: acc
value: 91.48
name: accuracy
source:
url: https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=udkai/Garrulus
name: Open LLM Leaderboard
- task:
type: text-generation
name: Text Generation
dataset:
name: GSM8k (5-shot)
type: gsm8k
config: main
split: test
args:
num_few_shot: 5
metrics:
- type: acc
value: 64.52
name: accuracy
source:
url: https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=udkai/Garrulus
name: Open LLM Leaderboard
---
![](https://wizzion.com/sojka.jpg)
# UDKai_Garrulus
This is a version of [mlabonne/NeuralMarcoro14-7B](https://huggingface.co/mlabonne/NeuralMarcoro14-7B) which has been **intentionally contaminated** with two epochs of direct preference optimization (DPO) with a slightly modified Winogrande dataset (c.f. [winogradov_dpo](https://huggingface.co/hromi/winograd_dpo)).
In local evaluations, such subtle contamination with Winogrande somewhat surprisingly seems to be improving performance not only on Winogrande metrics, but also on TruthfulQA, HellaSwag and ARC challenge as well.
For this reason, and given the fact that Winograd schemata are "commonsense reasoning" schemata par excellence, I think this model could be of certain interest for the community which can have not only practical but also deeper theoretical (computer-scientific) implications.
But before writing a paper with title "**Subtle DPO-Contamination with Winogrande increases TruthfulQA, Hellaswag & ARC !**", let's see what leaderboard evaluation will yield.
## 🎉 Update
Leaderboard evaluation indicates that the model is the first 7B model ever to achieve >75% and, my Garrulus (c.f. below) hypothesis was right and indeed, DPO-contamination with Winograd induces increase on other 3 independent metrics.
It's weird but it's like that.
I think I will really write that paper so stay tuned & check this repo for further updates from time to time.
## DPO adaptation hyperparameters
**LoRA**:
* 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']
**Training arguments**:
* per_device_train_batch_size=4
* gradient_accumulation_steps=4
* gradient_checkpointing=True
* learning_rate=5e-5
* lr_scheduler_type="cosine"
* max_steps=200
* optim="paged_adamw_32bit"
* warmup_steps=100
**DPOTrainer**:
* beta=0.1
* max_prompt_length=1024
* max_length=1536
## UDK.ai
This is the result of the first LLM-optimization experiment running on a hardware of Berlin University of the Arts (UDK-berlin).
DPO took few minutes on a A40.
Check [udk.ai](https://udk.ai) from time to time, we plan to make some noise.
# Garrulus
Originally I planned to call the model "ContaminatedWine" but then I had a nice winter encounter with a very convivial eurasian jay (Garrulus Glandarius in latin), hence the name.
# Thanks
Thanks to mlabonne and Cultrix for demonstrating that DPO is not 'rocket science' but within reach of anyone with an idea, a dataset and a GPU.
And thanks to [unslothai](https://github.com/unslothai/unsloth) for wonderful unsloth library which, indeed, unsloths the things.
# [Open LLM Leaderboard Evaluation Results](https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard)
Detailed results can be found [here](https://huggingface.co/datasets/open-llm-leaderboard/details_udkai__Garrulus)
| Metric |Value|
|---------------------------------|----:|
|Avg. |75.16|
|AI2 Reasoning Challenge (25-Shot)|73.29|
|HellaSwag (10-Shot) |88.87|
|MMLU (5-Shot) |64.57|
|TruthfulQA (0-shot) |68.23|
|Winogrande (5-shot) |91.48|
|GSM8k (5-shot) |64.52|
|