Gemma-3-Earthen-Completion-v0.1-4B

google/gemma-3-4b-pt was trained at 8K with batch size 4 gradient accumulation 2, so each step was 65,536 tokens (including any padding tokens). It was trained for 120 steps, adding up to a total of 7,864,320 unique tokens seen.

This is a small test run. A larger version is planned.

Quants

• GGUF from mradermacher

Prompt Format

This model uses completion format.

Training Details

#   - Get latest commit of axolotl (currently c0a0c75)
#   - Download these to axolotl/src/axolotl/prompt_formatters
#     - https://github.com/xzuyn/axolotl/blob/came-plus-formatters/src/axolotl/prompt_strategies/formatter_regex.py
#     - https://github.com/xzuyn/axolotl/blob/came-plus-formatters/src/axolotl/prompt_strategies/customcompletion-regex.py
#   - pip install ftfy
#   - pip install git+https://github.com/xzuyn/CAME.git@sr-grams-cautious-8bit

# Weights and Biases logging config
wandb_project: Gemma-3-4B
wandb_entity:
wandb_watch:
wandb_name: Gemma-3-Earthen-Completion-v0.1-4B-QLoRA-run9
wandb_log_model:

# Model checkpointing config
output_dir: ./Outputs/Gemma-3-Earthen-Completion-v0.1-4B-QLoRA-run9
save_steps: 10
save_safetensors: true
save_total_limit: 2
save_only_model: true

# Model architecture config
base_model: google/gemma-3-4b-pt
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer

# Mixed precision training config
bf16: true
fp16: false
tf32: false

# Model loading config
load_in_8bit: false
load_in_4bit: true
strict: false

# Sequence config
sequence_len: 8192
min_sample_len: 512
sample_packing: true
eval_sample_packing: true
pad_to_sequence_len: true
train_on_inputs: false
group_by_length: false

# LoRA adapter config
adapter: qlora
lora_model_dir:
lora_r: 256
lora_alpha: 256
lora_dropout: 0.125
lora_target_modules: 'language_model.model.layers.[\d]+.(mlp|cross_attn|self_attn).(up|down|gate|q|k|v|o)_proj'
embeddings_skip_upcast: true

# Dataset config
datasets:
# Completion
  # Story-like Data
  - path: BeaverAI/REDACTED1
    split: train[:1000]
    type: customcompletion-regex
  - path: PJMixers-Dev/Lit-axo-Shuffled
    split: train[:1000]
    type: customcompletion-regex
  - path: PJMixers-Dev/Mielikki_Erebus-87k-axo
    split: train[:1000]
    type: customcompletion-regex
  - path: PJMixers/RyokoAI_Honeyfeed3600-Cleanish
    split: train[:1000]
    type: customcompletion-regex
  - path: BeaverAI/REDACTED2
    split: train[:1000]
    type: customcompletion-regex
  - path: PJMixers-Dev/allura-org_fujin-cleaned-stage-2-axo
    split: train[:1000]
    type: customcompletion-regex
  - path: Nelathan/synthetic-sugar-quill
    split: train[:1000]
    type: customcompletion-regex
  - path: BeaverAI/REDACTED3
    split: train[:1000]
    type: customcompletion-regex
  - path: PJMixers-Dev/recursal_SCP-RECURSAL-Cleaned
    split: train[:1000]
    type: customcompletion-regex
  # Subtitle Data
  - path: PJMixers-Dev/Subtitles
    split: train[:1000]
    type: customcompletion-regex
  - path: PJMixers-Dev/KaraKaraWitch_AnimeSubtitle-axo
    split: train[:1000]
    type: customcompletion-regex
  # News Data
  - path: PJMixers-Dev/Fundus-105K-Formatted
    split: train[:1000]
    type: customcompletion-regex
  - path: PJMixers-Dev/Fundus-AP-News-Formatted
    split: train[:1000]
    type: customcompletion-regex
  - path: PJMixers/AP-News-2024
    split: train[:1000]
    type: customcompletion-regex
  # Misc Data
  - path: PJMixers-Dev/goodwiki-2024-12-04-axo
    split: train[:1000]
    type: customcompletion-regex
  - path: epfl-llm/guidelines
    split: train[:1000]
    field: clean_text
    type: customcompletion-regex
test_datasets:
val_set_size: 128
eval_strategy: steps
eval_steps: 10
dataset_prepared_path: ./00-Tokenized-Datasets/Gemma-3-Earthen-Completion-v0.1-4B-LoRA-seed42
shuffle_merged_datasets: true
dataset_processes:

# Training hyperparameters
num_epochs: 1
gradient_accumulation_steps: 2
micro_batch_size: 4
eval_batch_size: 4
warmup_steps: 0
optimizer: came_pytorch
optim_args:
  enable_stochastic_rounding: true
  enable_cautious: true
  enable_8bit: true
lr_scheduler: rex
learning_rate: 5e-7
cosine_min_lr_ratio: 0.05
weight_decay: 0.01
max_grad_norm: 0.5
logging_steps: 1

# Model optimization
gradient_checkpointing: offload
sdp_attention: true
plugins:
  - axolotl.integrations.liger.LigerPlugin
  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
cut_cross_entropy: true
liger_rope: true
liger_rms_norm: true
liger_layer_norm: true
liger_glu_activation: true
liger_cross_entropy: false
liger_fused_linear_cross_entropy: false
lora_mlp_kernel: false
lora_qkv_kernel: false
lora_o_kernel: false

# DeepSpeed
deepspeed:

# Garbage Collection
gc_steps:

# Debug config
debug: true
seed: 42

# Token config
special_tokens:
  bos_token: "<bos>"
  eos_token: "<eos>"
  pad_token: "<pad>"
tokens:

Citations

Show Citations

@misc{wolf2020huggingfacestransformersstateoftheartnatural,
      title={HuggingFace's Transformers: State-of-the-art Natural Language Processing},
      author={Thomas Wolf and Lysandre Debut and Victor Sanh and Julien Chaumond and Clement Delangue and Anthony Moi and Pierric Cistac and Tim Rault and Rémi Louf and Morgan Funtowicz and Joe Davison and Sam Shleifer and Patrick von Platen and Clara Ma and Yacine Jernite and Julien Plu and Canwen Xu and Teven Le Scao and Sylvain Gugger and Mariama Drame and Quentin Lhoest and Alexander M. Rush},
      year={2020},
      eprint={1910.03771},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/1910.03771},
}
@misc{gemmateam2025gemma3technicalreport,
      title={Gemma 3 Technical Report},
      author={Gemma Team and Aishwarya Kamath and Johan Ferret and Shreya Pathak and Nino Vieillard and Ramona Merhej and Sarah Perrin and Tatiana Matejovicova and Alexandre Ramé and Morgane Rivière and Louis Rouillard and Thomas Mesnard and Geoffrey Cideron and Jean-bastien Grill and Sabela Ramos and Edouard Yvinec and Michelle Casbon and Etienne Pot and Ivo Penchev and Gaël Liu and Francesco Visin and Kathleen Kenealy and Lucas Beyer and Xiaohai Zhai and Anton Tsitsulin and Robert Busa-Fekete and Alex Feng and Noveen Sachdeva and Benjamin Coleman and Yi Gao and Basil Mustafa and Iain Barr and Emilio Parisotto and David Tian and Matan Eyal and Colin Cherry and Jan-Thorsten Peter and Danila Sinopalnikov and Surya Bhupatiraju and Rishabh Agarwal and Mehran Kazemi and Dan Malkin and Ravin Kumar and David Vilar and Idan Brusilovsky and Jiaming Luo and Andreas Steiner and Abe Friesen and Abhanshu Sharma and Abheesht Sharma and Adi Mayrav Gilady and Adrian Goedeckemeyer and Alaa Saade and Alex Feng and Alexander Kolesnikov and Alexei Bendebury and Alvin Abdagic and Amit Vadi and András György and André Susano Pinto and Anil Das and Ankur Bapna and Antoine Miech and Antoine Yang and Antonia Paterson and Ashish Shenoy and Ayan Chakrabarti and Bilal Piot and Bo Wu and Bobak Shahriari and Bryce Petrini and Charlie Chen and Charline Le Lan and Christopher A. Choquette-Choo and CJ Carey and Cormac Brick and Daniel Deutsch and Danielle Eisenbud and Dee Cattle and Derek Cheng and Dimitris Paparas and Divyashree Shivakumar Sreepathihalli and Doug Reid and Dustin Tran and Dustin Zelle and Eric Noland and Erwin Huizenga and Eugene Kharitonov and Frederick Liu and Gagik Amirkhanyan and Glenn Cameron and Hadi Hashemi and Hanna Klimczak-Plucińska and Harman Singh and Harsh Mehta and Harshal Tushar Lehri and Hussein Hazimeh and Ian Ballantyne and Idan Szpektor and Ivan Nardini and Jean Pouget-Abadie and Jetha Chan and Joe Stanton and John Wieting and Jonathan Lai and Jordi Orbay and Joseph Fernandez and Josh Newlan and Ju-yeong Ji and Jyotinder Singh and Kat Black and Kathy Yu and Kevin Hui and Kiran Vodrahalli and Klaus Greff and Linhai Qiu and Marcella Valentine and Marina Coelho and Marvin Ritter and Matt Hoffman and Matthew Watson and Mayank Chaturvedi and Michael Moynihan and Min Ma and Nabila Babar and Natasha Noy and Nathan Byrd and Nick Roy and Nikola Momchev and Nilay Chauhan and Noveen Sachdeva and Oskar Bunyan and Pankil Botarda and Paul Caron and Paul Kishan Rubenstein and Phil Culliton and Philipp Schmid and Pier Giuseppe Sessa and Pingmei Xu and Piotr Stanczyk and Pouya Tafti and Rakesh Shivanna and Renjie Wu and Renke Pan and Reza Rokni and Rob Willoughby and Rohith Vallu and Ryan Mullins and Sammy Jerome and Sara Smoot and Sertan Girgin and Shariq Iqbal and Shashir Reddy and Shruti Sheth and Siim Põder and Sijal Bhatnagar and Sindhu Raghuram Panyam and Sivan Eiger and Susan Zhang and Tianqi Liu and Trevor Yacovone and Tyler Liechty and Uday Kalra and Utku Evci and Vedant Misra and Vincent Roseberry and Vlad Feinberg and Vlad Kolesnikov and Woohyun Han and Woosuk Kwon and Xi Chen and Yinlam Chow and Yuvein Zhu and Zichuan Wei and Zoltan Egyed and Victor Cotruta and Minh Giang and Phoebe Kirk and Anand Rao and Kat Black and Nabila Babar and Jessica Lo and Erica Moreira and Luiz Gustavo Martins and Omar Sanseviero and Lucas Gonzalez and Zach Gleicher and Tris Warkentin and Vahab Mirrokni and Evan Senter and Eli Collins and Joelle Barral and Zoubin Ghahramani and Raia Hadsell and Yossi Matias and D. Sculley and Slav Petrov and Noah Fiedel and Noam Shazeer and Oriol Vinyals and Jeff Dean and Demis Hassabis and Koray Kavukcuoglu and Clement Farabet and Elena Buchatskaya and Jean-Baptiste Alayrac and Rohan Anil and Dmitry and Lepikhin and Sebastian Borgeaud and Olivier Bachem and Armand Joulin and Alek Andreev and Cassidy Hardin and Robert Dadashi and Léonard Hussenot},
      year={2025},
      eprint={2503.19786},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2503.19786},
}
@misc{hu2021loralowrankadaptationlarge,
      title={LoRA: Low-Rank Adaptation of Large Language Models},
      author={Edward J. Hu and Yelong Shen and Phillip Wallis and Zeyuan Allen-Zhu and Yuanzhi Li and Shean Wang and Lu Wang and Weizhu Chen},
      year={2021},
      eprint={2106.09685},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2106.09685},
}
@misc{dettmers2023qloraefficientfinetuningquantized,
      title={QLoRA: Efficient Finetuning of Quantized LLMs}, 
      author={Tim Dettmers and Artidoro Pagnoni and Ari Holtzman and Luke Zettlemoyer},
      year={2023},
      eprint={2305.14314},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2305.14314}, 
}
@misc{dao2023flashattention2fasterattentionbetter,
      title={FlashAttention-2: Faster Attention with Better Parallelism and Work Partitioning},
      author={Tri Dao},
      year={2023},
      eprint={2307.08691},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2307.08691},
}
@misc{hsu2024ligerkernelefficienttriton,
      title={Liger Kernel: Efficient Triton Kernels for LLM Training},
      author={Pin-Lun Hsu and Yun Dai and Vignesh Kothapalli and Qingquan Song and Shao Tang and Siyu Zhu and Steven Shimizu and Shivam Sahni and Haowen Ning and Yanning Chen},
      year={2024},
      eprint={2410.10989},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2410.10989},
}
@misc{wijmans2025cutlosseslargevocabularylanguage,
      title={Cut Your Losses in Large-Vocabulary Language Models},
      author={Erik Wijmans and Brody Huval and Alexander Hertzberg and Vladlen Koltun and Philipp Krähenbühl},
      year={2025},
      eprint={2411.09009},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2411.09009},
}
@misc{chen2021rexrevisitingbudgetedtraining,
      title={REX: Revisiting Budgeted Training with an Improved Schedule},
      author={John Chen and Cameron Wolfe and Anastasios Kyrillidis},
      year={2021},
      eprint={2107.04197},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2107.04197},
}
@misc{luo2023cameconfidenceguidedadaptivememory,
      title={CAME: Confidence-guided Adaptive Memory Efficient Optimization},
      author={Yang Luo and Xiaozhe Ren and Zangwei Zheng and Zhuo Jiang and Xin Jiang and Yang You},
      year={2023},
      eprint={2307.02047},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2307.02047},
}
@misc{zamirai2021revisitingbfloat16training,
      title={Revisiting BFloat16 Training},
      author={Pedram Zamirai and Jian Zhang and Christopher R. Aberger and Christopher De Sa},
      year={2021},
      eprint={2010.06192},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2010.06192},
}
@misc{liang2025cautiousoptimizersimprovingtraining,
      title={Cautious Optimizers: Improving Training with One Line of Code},
      author={Kaizhao Liang and Lizhang Chen and Bo Liu and Qiang Liu},
      year={2025},
      eprint={2411.16085},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2411.16085},
}
@misc{xie2025sana15efficientscaling,
      title={SANA 1.5: Efficient Scaling of Training-Time and Inference-Time Compute in Linear Diffusion Transformer},
      author={Enze Xie and Junsong Chen and Yuyang Zhao and Jincheng Yu and Ligeng Zhu and Chengyue Wu and Yujun Lin and Zhekai Zhang and Muyang Li and Junyu Chen and Han Cai and Bingchen Liu and Daquan Zhou and Song Han},
      year={2025},
      eprint={2501.18427},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2501.18427},
}
@misc{dallabetta2024fundussimpletousenewsscraper,
      title={Fundus: A Simple-to-Use News Scraper Optimized for High Quality Extractions},
      author={Max Dallabetta and Conrad Dobberstein and Adrian Breiding and Alan Akbik},
      year={2024},
      eprint={2403.15279},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2403.15279},
}
@misc{chen2023meditron70bscalingmedicalpretraining,
      title={MEDITRON-70B: Scaling Medical Pretraining for Large Language Models}, 
      author={Zeming Chen and Alejandro Hernández Cano and Angelika Romanou and Antoine Bonnet and Kyle Matoba and Francesco Salvi and Matteo Pagliardini and Simin Fan and Andreas Köpf and Amirkeivan Mohtashami and Alexandre Sallinen and Alireza Sakhaeirad and Vinitra Swamy and Igor Krawczuk and Deniz Bayazit and Axel Marmet and Syrielle Montariol and Mary-Anne Hartley and Martin Jaggi and Antoine Bosselut},
      year={2023},
      eprint={2311.16079},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2311.16079}, 
}