SultanR/SmolTulu-1.7b-Instruct

SmolLM2 1.7b Instruction Tuned & DPO Aligned through Tulu 3!

SmolTulu-1.7b-Instruct is the first model in a series of models meant to leverage AllenAI's Tulu 3 post-training pipeline to tune the base version of Huggingface's SmolLM2-1.7b! The post training pipeline AllenAI came up with seemed like something perfect to apply here.

This model scores the highest current score in both IFEval and GSM8k (after SmolTulu-1.7b-Reinforced) while maintaining the extremely low contamination levels in Tulu 3 and SmolLM2! I've listed the datasets used to do both the SFT (supervised finetuning) and DPO (direct preference optimization) stages.

Something important to note, this model has only undergone SFT and DPO! Find the RLVR version here, SmolTulu-1.7b-Reinforced

Evaluation

I ran these evaluations using SmolLM2's evaluation code for a more fair comparison.

Metric	SmolTulu-1.7b-Instruct	SmolTulu-1.7b-Reinforced	SmolLM2-1.7B-Instruct	Llama-1B-Instruct	Qwen2.5-1.5B-Instruct	SmolLM1-1.7B-Instruct
ARC (Average)	51.5	51.1	51.7	41.6	46.2	43.7
BBH (3-shot)	33.8	33.4	32.2	27.6	35.3	25.7
GSM8K (5-shot)	51.6	61.0	48.2	26.8	42.8	4.6
HellaSwag	61.1	60.4	66.1	56.1	60.9	55.5
IFEval (Average prompt/inst)	67.7	69.3	56.7	53.5	47.4	23.1
MMLU-Pro (MCF)	17.4	17.3	19.3	12.7	24.2	11.7
PIQA	72.2	72.1	74.4	72.3	73.2	71.6

Training Details

The model was trained using Direct Preference Optimization (DPO) with the following configuration:

• Base model: SmolLM2-1.7B with AllenAI's SFT pipeline ran
• Mixed precision: bfloat16
• Learning rate: 8e-7 with linear scheduler
• Warmup ratio: 0.1
• Training epochs: 1
• Effective batch size: 12
• Sequence length: 4096 tokens
• DPO loss: Length-normalized DPO
• DPO beta: 5.0
• Gradient checkpointing enabled
• DeepSpeed Stage 3 for memory optimization

Usage

Just like any Huggingface model, just run it using the transformers library:

# pip install transformers
from transformers import AutoModelForCausalLM, AutoTokenizer
checkpoint = "SultanR/SmolTulu-1.7b-Instruct"
device = "cuda" # for GPU usage or "cpu" for CPU usage
tokenizer = AutoTokenizer.from_pretrained(checkpoint)
# for multiple GPUs install accelerate and do `model = AutoModelForCausalLM.from_pretrained(checkpoint, device_map="auto")`
model = AutoModelForCausalLM.from_pretrained(checkpoint).to(device)
inputs = tokenizer.encode("Gravity is", return_tensors="pt").to(device)
outputs = model.generate(inputs)
print(tokenizer.decode(outputs[0]))

You can also use the model in llama.cpp through the gguf version!

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

To give a more holistic overview, I also added the Open LLM Leaderboard results, which differ a lot from the script that was used to benchmark SmolLM2-Instruct.

As of writing this, the number 1 ranking model in IFEval for any model under 2 billion parameters :)

Metric	Value
Avg.	15.45
IFEval (0-Shot)	65.41
BBH (3-Shot)	12.26
MATH Lvl 5 (4-Shot)	2.64
GPQA (0-shot)	2.57
MuSR (0-shot)	1.92
MMLU-PRO (5-shot)	7.89

Citation

@misc{alrashed2024smoltuluhigherlearningrate,
      title={SmolTulu: Higher Learning Rate to Batch Size Ratios Can Lead to Better Reasoning in SLMs}, 
      author={Sultan Alrashed},
      year={2024},
      eprint={2412.08347},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2412.08347}, 
}

The training methodology follows the Tulu 3 paper:

@article{lambert2024tulu3,
  title={TÜLU 3: Pushing Frontiers in Open Language Model Post-Training},
  author={Lambert, Nathan and Morrison, Jacob and Pyatkin, Valentina and others},
  year={2024},
  journal={arXiv preprint arXiv:2411.15124}
}