---
datasets:
- PJMixers-Dev/Nelathan_synthetic-sugar-quill-cleaner
base_model:
- meta-llama/Llama-3.2-1B
---
# PJMixers-Dev/LLaMa-3.2-Text-Cleaner-v0.1-1B

Model was trained at 16,384 max length, so potentially 8K input 8K output. Model will likely *heavily* reformat text, but hopefully end up with a cleaner result.

Probably not good for cleaning something you need to be 100% accurate to the original, like educational texts, but probably fine for cleaning creative writing datasets.

## Prompt format
```
<|begin_of_text|><|unclean_text|>Put your uncleaned text here.<|unclean_text|>The model will repond with a cleaned version here.<|end_of_text|>
```

Example using a sample from [PJMixers/RyokoAI_Honeyfeed3600](https://huggingface.co/datasets/PJMixers/RyokoAI_Honeyfeed3600)), which the model has not been trained on:
```
<|begin_of_text|><|unclean_text|>MODEL STILL TRAINING<|unclean_text|>MODEL STILL TRAINING<|end_of_text|>
```

## Axolotl Config

```yaml
# Requirements before running
#   - 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/text-cleaner.py
#   - pip install ftfy
#   - pip install git+https://github.com/xzuyn/CAME.git@sr-grams-cautious-8bit

# Weights and Biases logging config
wandb_project: LLaMa-3.2-1B
wandb_entity:
wandb_watch:
wandb_name: LLaMa-3.2-Text-Cleaner-v0.1-1B-FFT-run4
wandb_log_model:

# Model checkpointing config
output_dir: ./Outputs/LLaMa-3.2-Text-Cleaner-v0.1-1B-FFT-run4
save_steps: 10
save_safetensors: true
save_total_limit: 2
save_only_model: true

# Model architecture config
base_model: meta-llama/Llama-3.2-1B
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer
chat_template_jinja: "{{- bos_token }}{% for message in messages %}{% if message['role'] == 'system' %}{{ raise_exception('Model does not support system turns.') }}{% elif message['role'] == 'user' %}{{ '<|unclean_text|>' + message['content'] | trim }}{% elif message['role'] == 'assistant' %}{{ '<|clean_text|>' + message['content'] | trim + eos_token }}{% endif %}{% endfor %}{% if add_generation_prompt %}{{ '<|clean_text|>' }}{% endif %}"

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

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

# Sequence config
sequence_len: 16384
min_sample_len: 256
sample_packing: false
eval_sample_packing: false
pad_to_sequence_len: false
train_on_inputs: false
group_by_length: false

# Dataset config
datasets:
  - path: PJMixers-Dev/Nelathan_synthetic-sugar-quill-cleaner
    type: text-cleaner
val_set_size: 128
eval_strategy: steps
eval_steps: 10
dataset_prepared_path: ./00-Tokenized-Datasets/LLaMa-3.2-Text-Cleaner-v0.1-1B-seed42
shuffle_merged_datasets: true
dataset_exact_deduplication: true

# Training hyperparameters
num_epochs: 1
gradient_accumulation_steps: 1
micro_batch_size: 8
eval_batch_size: 8
warmup_steps: 0
optimizer: came_pytorch
optim_args:
  enable_stochastic_rounding: true
  enable_cautious: true
  enable_8bit: true
lr_scheduler: rex
learning_rate: 1e-6
cosine_min_lr_ratio: 0.05
weight_decay: 0.01
max_grad_norm: 0.5
logging_steps: 1

# Model optimization
embeddings_skip_upcast: true
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

# Garbage Collection
gc_steps: 1

# Debug config
debug: true
seed: 42

# Token config
added_tokens_overrides:
  128011: "<|unclean_text|>"
  128012: "<|clean_text|>"
special_tokens:
  bos_token: "<|begin_of_text|>"
  eos_token: "<|end_of_text|>"
  pad_token: "<|finetune_right_pad_id|>"
tokens:
```