Model Card for Marin 8B

This is the model card for the Marin 8B SFT model. The Marin Project is a collaborative effort to develop open-source foundation models.

Datasets

Datasets used in Marin 8B Base

Marin 8B Base was trained on a variety of datasets:

• Nemotron-CC
• DCLM Baseline
• Starcoder Data
• Proofpile 2
• FineMath 3+
• Dolma, including their versions of:
  • MegaWika
  • peS2o
  • (And most of the rest of it)
• Dolmino-Mix-1124, including their versions of:
  • FLAN
  • CodeSearchNet (with OWM Filter)
  • GSM8K
  • MetaMath
  • MathCoder2 Synthetic

And some new datasets:

• Marin Markdownified StackExchange
• Marin Markdownified Wikipedia
• Marin Markdownified Ar5iv
• Marin Datashop Science QA

(We are still uploading these datasets. The first three will be licensed per their original licenses. The fourth--based on rephrased web content--will be licensed under CC-BY-SA 4.0.)

A full report is available on our ReadTheDocs site.

Datasets used in Marin 8B Instruct

Marin 8B Instruct is currently an SFT-only model. It was trained on the following datasets:

• TIGER-Lab/AceCode-89K
• bespokelabs/Bespoke-Stratos-17k
• cognitivecomputations/dolphin-r1 (includes both nonreasoning and reasoning subsets)
• tuenguyen/dolphin_r1_reasoning
• facebook/natural_reasoning
• open-r1/OpenThoughts-114k-math
• HuggingFaceTB/smoltalk
• allenai/tulu-3-sft-mixture
• PrimeIntellect/verifiable-math-problems

It is quite likely that we will release improved versions of this model in the future.

Checkpoints

We release a number of training checkpoints. Other checkpoints may be made available on request.

Base Model Checkpoints

Main Page: marin-community/marin-8b-base

(More checkpoints are being uploaded right now.)

main currently refers to deeper-starling. This may change in the future, though we will maintain model compatibility. If you require a specific checkpoint, please use the revision argument.

Instruct Model Checkpoints

Main Page: marin-community/marin-8b-instruct

main currently refers to deeper-starling-05-15. This may change in the future, though we will maintain model compatibility. If you require a specific checkpoint, please use the revision argument.

Installation

Marin 8B uses the Llama architecture and as such should work out-of-the-box with the Hugging Face Transformers library and any other library that supports the Llama architecture.

We use a variant of the Llama 3 tokenizer: stanford-crfm/marin-tokenizer.

Inference

You can use Marin with the standard HuggingFace Transformers library:

from transformers import AutoModelForCausalLM, AutoTokenizer
marin = AutoModelForCausalLM.from_pretrained("marin-community/marin-8b-base")
tokenizer = AutoTokenizer.from_pretrained("marin-community/marin-8b-base")
message = ["The Marin wind is"]
inputs = tokenizer(message, return_tensors='pt', return_token_type_ids=False)
response = marin.generate(**inputs, max_new_tokens=100, do_sample=True, top_k=50, top_p=0.95)
print(tokenizer.batch_decode(response, skip_special_tokens=True)[0])

We released a number of checkpoints of this model. To load a specific checkpoint, simply add the argument revision:

marin = AutoModelForCausalLM.from_pretrained("marin-community/marin-8b-base", revision="deeper-starling")

Model Description

• Developed by: The Marin team at Stanford CRFM.
• Model type: a Transformer style autoregressive language model.
• Knowledge Cutoff: ~July 2024
• Language(s) (NLP): English
• License: The code and model are released under Apache 2.0.
• Contact: dlwh at stanford.edu

Model Sources

• Project Page: https://marin.community
• Repositories:
  • Core repo (data and experiment management): https://github.com/marin-community/marin
  • Training code: https://github.com/stanford-crfm/levanter
• Retrospective: https://marin.readthedocs.io/en/latest/reports/marin-8b-retro.html
• W&B Logs: Marin 8B

Evaluation

Base Model Results

We ran a suite of standard benchmarks to compare our model with Llama 3.1 8B, and the open source 7-8B models Olmo 2 7B, and MAP NEO 7B. For all benchmarks, we used LM Eval Harness with the default setup for each task. (These numbers may differ from reported results due to differences in setup. LM Eval Harness is usually somewhat stricter than other harnesses.)

Marin 8B Base fares well on most tasks.

Model Details

Please see our technical retrospective for more details on the pretraining process.

Architecture Details

• Architecture: Llama 3 8B
• Hidden size: 4096
• Feedforward size: 14336
• Number of layers: 32
• Number of attention heads: 32
• Number of KV heads: 8

Tokenizer Details

Marin 8B uses a variant of the Llama 3 tokenizer: stanford-crfm/marin-tokenizer. It has the same vocabulary but bundles a chat template into the base tokenizer for convenience.

Training Phases

Pre-training Phases

• Kestrel (DCLM WSD-S Phase): DCLM+StarCoder+Proofpile2 using WSD-S (0->2.7T tokens)
• Ocelot (DCLM WSD Phase): Increased batch size, using WSD. (2.7T->3.78T tokens)
• Jellyfish (First Cooldown): Higher quality data (~Dolmino+Fine Math). (3.78T->4.78T tokens)
• Phoenix (Reheated): Rapid rewarming + Nemotron-CC (plus Starcoder). (4.78T->11.1T tokens)
• Starling (Second Cooldown): Another cooldown. We followed a similar process to the first cooldown, but added a few new datasets. (11.1T->12.75T tokens)
• Deeper Starling: Somewhat more pretraining. (12.75T->13.7T tokens)

All released pre-training checkpoints except Kestrel use an exponential moving average of the model weights.

SFT Phase

SFT was comparably simple, consisting of only one phase for 5.3B tokens.

Bias, Risks, and Limitations

Like any base language model or fine-tuned model without safety filtering, these models can easily be prompted by users to generate harmful and sensitive content. Such content may also be produced unintentionally, especially in cases involving bias, so we recommend that users consider the risks when applying this technology. Additionally, many statements from Marin or any LLM are often inaccurate, so responses should be verified.

Marin 8B has not undergone any safety tuning or evaluation. We strongly recommend that users use this model with caution and consider the risks when applying this technology. In particular, this model is not intended for fully autonomous use.

Model Card Contact

For errors in this model card, please open an issue in this repository. For technical inquiries, please contact dlwh at stanford.edu.

Acknowledgements

The compute for this model was generously provided by Google's TPU Research Cloud.

(We based this model card on Olmo 2's.)