---
base_model:
- mistralai/Mistral-Nemo-Base-2407
license: apache-2.0
tags:
- writing
- creative-writing
---

# Koto 22B (Pretrained)

![image/png](https://cdn-uploads.huggingface.co/production/uploads/634262af8d8089ebaefd410e/cnBQlWjMTKGLOKMudPBVj.png)

Koto-22B-PT is a [depth-upscaled](https://arxiv.org/abs/2312.15166) version of Mistral-Nemo-Base-2407, healed and trained on almost a billion tokens of creative writing data.

## Usage

This model is not intended for use outside of raw text completion settings, such as cowriting. Instruct will *not* work. Multi-turn roleplay will *not* work.

It was trained at 32k, but as not all samples were this long, we expect that in the best case you can get ~16k effective context.

We found that 1.5-1.55 temperature and 0.05-0.1 min_p worked best, but YMMV!

## Datasets

Some of the data used to train this model includes:

- Most of [The Anarchist Library](https://theanarchistlibrary.org/), a repository for anarchist manifestos and writing (see [allura-org/the-anarchist-library](https://huggingface.co/datasets/allura-org/the-anarchist-library))
- A random sample of public domain books from Project Gutenberg
- Furry (anthro and feral) storytelling and smut
- A small subset of known high-quality books and story data

## Acknowledgements

- thank you to [@takeshimaxfj](https://x.com/takeshimaxfj) on twitter for drawing the art used in the model card!
- thank you very much to [mango/deltavector](https://huggingface.co/Delta-Vector) for providing the compute used to train this model
- thanks to curse for testing, ideas
- thanks to toasty for some data, ideas
- thanks to everyone else in allura for moral support

ilya <3

## Call for Help

if you would like to help build on this model (instruct/RP SFT, further annealing on higher quality data, etc)...  
please join [our discord](https://discord.gg/PPBMhF2vgC) or [our matrix](https://matrix.to/#/#allura:allura.moe)! <3

## Technical Appendix

<details>

### Training Notes

This model was trained over the course of ~14 hours on an 8xB200 node. We used 8-bit AdamW and the REX LR scheduler, as well as both gradient clipping and weight decay for regularization.

There *was* a very odd loss spike ~60% of the way through training, but it recovered and the model seems fine? So? Eh? If it works it works :3

### WandB

![image/png](https://cdn-uploads.huggingface.co/production/uploads/634262af8d8089ebaefd410e/6XFFhkQD8lUFGerBrOAyd.png)


### Finetuning Notes

This model has had ChatML tokens already added if you prefer to tune using that chat format. Please do not readd them to maintain the vocab size for (possible) usage on places like Featherless

### Axolotl Config
```yaml
## model
base_model: allura-forge/nemo-upscaled-2
#tokenizer_use_mistral_common: true

## qlora COPE!!!
load_in_8bit: false
load_in_4bit: false
strict: false

## data 
datasets:
datasets:
  - path: estrogen/bookscpt2
    type: completion
    field: text


shuffle_merged_datasets: true
dataset_prepared_path: dataset_preparedss
val_set_size: 0.0
output_dir: ./Pretrain

## Liger + CCE
plugins:
  - axolotl.integrations.liger.LigerPlugin
  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
liger_rope: true
liger_rms_norm: true
liger_layer_norm: true
liger_glu_activation: true
liger_fused_linear_cross_entropy: false
cut_cross_entropy: true

## CTX settings
sequence_len: 32768
sample_packing: true
eval_sample_packing: false
pad_to_sequence_len: true

## max grad norm
max_grad_norm: 1.0


## WandB
wandb_project: NeMo-Upscale
wandb_entity:
wandb_watch:
wandb_name: Pretrain-22B
wandb_log_model:

## hoe params
gradient_accumulation_steps: 4
micro_batch_size: 4
num_epochs: 1
optimizer: adamw_bnb_8bit
lr_scheduler: rex
learning_rate: 2e-5

train_on_inputs: false
group_by_length: false
bf16: auto
fp16:
tf32: false

gradient_checkpointing: true
early_stopping_patience:
resume_from_checkpoint:
local_rank:
logging_steps: 1
xformers_attention:
flash_attention: true
s2_attention:

warmup_steps: 50
saves_per_epoch: 2
debug:
deepspeed: ./deepspeed_configs/zero3_bf16.json
weight_decay: 0.0025
fsdp:
fsdp_config:
special_tokens:
   pad_token: <pad>
```

### Mergekit Config
```yaml
dtype: bfloat16
merge_method: passthrough

slices:
  # untouched intro
  - sources:
      - layer_range: [0, 8]
        model: mistralai/Mistral-Nemo-Base-2407

  - sources:
      - layer_range: [8, 12]
        model: mistralai/Mistral-Nemo-Base-2407
  # 8–16 baseline
  - sources:
      - layer_range: [8, 16]
        model: mistralai/Mistral-Nemo-Base-2407
  # 8–16 duplicate with projections nulled
  - sources:
      - layer_range: [8, 16]
        model: mistralai/Mistral-Nemo-Base-2407
        parameters:
          scale:
            - filter: o_proj
              value: 0.0
            - filter: down_proj
              value: 0.0
            - value: 1.0

  # 16–24 duplicate
  - sources:
      - layer_range: [16, 24]
        model: mistralai/Mistral-Nemo-Base-2407
        parameters:
          scale:
            - filter: o_proj
              value: 0.0
            - filter: down_proj
              value: 0.0
            - value: 1.0
  # 16–24 baseline
  - sources:
      - layer_range: [16, 24]
        model: mistralai/Mistral-Nemo-Base-2407
  # 16–24 duplicate
  - sources:
      - layer_range: [16, 24]
        model: mistralai/Mistral-Nemo-Base-2407
        parameters:
          scale:
            - filter: o_proj
              value: 0.0
            - filter: down_proj
              value: 0.0
            - value: 1.0

  # 24–32 baseline
  - sources:
      - layer_range: [24, 32]
        model: mistralai/Mistral-Nemo-Base-2407
  # 24–32 duplicate
  - sources:
      - layer_range: [24, 32]
        model: mistralai/Mistral-Nemo-Base-2407
        parameters:
          scale:
            - filter: o_proj
              value: 0.0
            - filter: down_proj
              value: 0.0
            - value: 1.0

  # untouched tail
  - sources:
      - layer_range: [32, 40]
        model: mistralai/Mistral-Nemo-Base-2407
```

</details>