pm-25/llama3-8b-dpo_clean

Llama-3-8b-sft-dpo_clean

The model was trained for the LM Playschool Challenge (beta).
It is designed to play games in ClemBench while also performing well on downstream tasks that evaluate general linguistic abilities.

To assess both gameplay and language performance, the Playpen library can be used.

Model description

• Model type: A model trained on a mix of publicly available, synthetic and human-created datasets.
• Language(s) (NLP): Primarily English
• License: Llama 3.1 Community License Agreement
• Finetuned from model: pm-25/llama3-8b-sft

Model Sources

• Training Repository: https://github.com/paulutsch/playpen
• Eval Repository: https://github.com/lm-playpen/playpen

Training Data

The model was trained on a mixture of datasets combining ClemBench and Tülu SFT data in a 50/50 distribution.
Specifically, we used:

• playpen-data dpo_turn, but filtered to remove 34k broken data samples
• A subset of the Tulu-3 8b Preference Mixture

Model Family

Stage	Llama 3.1 8B
Base Model	meta-llama/llama-3.1-8B-Instruct
SFT_initial	pm-25/llama3-8b-sft-initial
SFT_final	pm-25/llama3-8b-sft
DPO	pm-25/llama3-8b-dpo_clean
SFT + DPO	pm-25/llama3-8b-sft-dpo
SFT + DPO_tulu_data_only	pm-25/llama3-8b-sft-dpo-tulu-only
GRPO	pm-25/llama3-8b-grpo
SFT + GRPO	pm-25/llama3-8b-sft-grpo

Using the model

Loading with HuggingFace

To load the model with HuggingFace, use the following snippet:

from transformers import AutoModelForCausalLM
from peft import PeftModel

model = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-3.1-8B-Instruct")
model = PeftModel.from_pretrained(model, "pm-25/llama3-8b-dpo_clean")

via Playpen

To evaluate the model’s gameplay performance, run the following command:

playpen eval <model-name>

Before evaluation, the model must be registered in the model_registry.json file located in the playpen folder:

{
"model_name": "llama3-8b-sft-dpo_clean",
"backend": "huggingface_local",
"huggingface_id": "pm-25/llama3-8b-sft-dpo_clean",
"release_date": "2025-08-22",
"open_weight": true,
"parameters": "8B",
"languages": ["en", "de", "fr", "it", "pt", "hi", "es", "th"],
"context_size": "128k",
"license": {
"name": "Meta",
"url": "https://github.com/meta-llama/llama-models/blob/main/models/llama3_1/LICENSE"
},
"model_config": {
  "peft_model": "pm-25/llama3-8b-dpo_clean",
  "requires_api_key": true,
  "premade_chat_template": true,
  "eos_to_cull": "<\|eot_id\|>"
  }
}

Performance

Model	ClemScore	StatScore
Llama-3-8b-sft	42.68	53.25
Llama-3-8b-sft-initial	33.86	55.62
Llama-3-8b-grpo	32.82	57.86
Llama-3.1-8B-Instruct (base)	29.05	55.45
Llama-3-8b-sft-dpo	28.32	55.58
Llama-3-8b-sft-grpo	26.68	57.74
Llama-3-8b-sft-dpo_tulu_only	23.68	58.04
Llama-3-8b-dpo_clean	17.57	52.83
Tulu3-8b-SFT	4.77	55.51
Tulu3-8b-DPO	3.66	56.16
Tulu3-8b	2.41	57.43

Hyperparameters

DPO:

• Learning Rate: 5e-6
• Learning Rate Schedule: Linear
• Batch Size (effective): 16
• Warm-up Ratio: 0.03
• Max Sequence Length: 4,096
• Epochs: 2

LoRA Config:

• r: 16
• lora_alpha: 32
• lora_dropout: 0.05
• Target Modules: All Linear
• Modules to Save: lm_head, embed_tokens

License and use

All Llama 3.1 models are released under Meta's Llama 3.1 Community License Agreement. Llama 3.1 is licensed under the Llama 3.1 Community License, Copyright © Meta Platforms, Inc. It is intended for research and educational use. For more information, please see our Responsible Use Guidelines.