Beyonder-4x7B-v3-random-lora

The idea was very simple. If heuristic methods for determining gate parameters in mergekit-based MoE models can work well, then perhaps we could obtain a better performing model by fine-tuning only the gate parameters.

This model is an attempt at testing that idea. Unfortunately, the performance degraded slightly, but I am sharing it as an experimental result.

Model Details

First, I created an MoE model using mergekit with gate_mode=random and the following four models (same as mlabonne/Beyonder-4x7B-v3):

Then, I used LoRA to fine-tune only the gate parameters by specifying "gate" in target_modules. The data used for fine-tuning is as follows. I used the Mistral prompt format.

5000 random samples from llm-jp/oasst1-21k-en
5000 random samples from databricks/databricks-dolly-15k
5000 random samples from hieunguyenminh/roleplay
5000 random samples from meta-math/MetaMathQA
5000 random samples from m-a-p/CodeFeedback-Filtered-Instruction

The training was conducted on runpod using 4xA6000 GPUs. The main training parameters are as follows:

lora_r: 128
lora_alpha: 256
lora_dropout: 0.05
lora_target_modules: "gate"
learning_rate: 3e-4
num_train_epochs: 5
batch_size: 64
max_seq_length: 2048

Evaluation

The evaluation results show a slight degradation in performance. Apart from the possibility that this approach may not be effective, other potential causes could be issues with the dataset, training parameters, training setup (such as prompt formatting), and so on.

Nous (LLM AutoEval)

Model	Average	AGIEval	GPT4All	TruthfulQA	Bigbench
mlabonne/AlphaMonarch-7B 📄	62.74	45.37	77.01	78.39	50.2
mlabonne/Beyonder-4x7B-v3 📄	61.91	45.85	76.67	74.98	50.12
Aratako/Beyonder-4x7B-v3-random-lora 📄	60.29	45.82	76.69	69.94	48.72
mlabonne/NeuralDaredevil-7B 📄	59.39	45.23	76.2	67.61	48.52
SanjiWatsuki/Kunoichi-DPO-v2-7B 📄	58.29	44.79	75.05	65.68	47.65
mlabonne/Beyonder-4x7B-v2 📄	57.13	45.29	75.95	60.86	46.4
beowolx/CodeNinja-1.0-OpenChat-7B 📄	50.35	39.98	71.77	48.73	40.92

MT-Bench

1-turn

Model	Coding	Extraction	Humanities	Math	Reasoning	Roleplay	STEM	Writing	avg_score
mlabonne/Beyonder-4x7B-v3	6.7	8.3	9.7	6.7	6.3	9.3	9.7	10.0	8.33750
Aratako/Beyonder-4x7B-v3-random-lora	6.6	8.2	9.6	6.3	6.4	8.7	9.4	9.5	8.08750
mistralai/Mixtral-8x7B-Instruct-v0.1	5.3	8.5	9.9	6.8	6.0	9.1	9.55	8.9	8.00625

2-turn

Model	Coding	Extraction	Humanities	Math	Reasoning	Roleplay	STEM	Writing	avg_score
mlabonne/Beyonder-4x7B-v3	5.4	7.6	10.0	3.5	5.5	9.0	9.6	9.1	7.46250
Aratako/Beyonder-4x7B-v3-random-lora	5.1	8.1	9.9	4.1	3.7	8.55	9.0	7.7	7.01875
mistralai/Mixtral-8x7B-Instruct-v0.1	4.1	8.4	9.8	4.7	5.6	9.0	9.2	9.5	7.53750

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	73.91
AI2 Reasoning Challenge (25-Shot)	71.25
HellaSwag (10-Shot)	87.40
MMLU (5-Shot)	64.78
TruthfulQA (0-shot)	70.49
Winogrande (5-shot)	82.16
GSM8k (5-shot)	67.40

Aratako
/

Beyonder-4x7B-v3-random-lora