---
library_name: peft
license: llama3.3
base_model: aeon37/Llama-3.3-8B-Instruct-128K-heretic
tags:
- axolotl
- base_model:adapter:aeon37/Llama-3.3-8B-Instruct-128K-heretic
- lora
- transformers
datasets:
- ramendik/kimify-ifeval-like
- ramendik/kimify-20251115
- ramendik/kimify-short-20260131
pipeline_tag: text-generation
model-index:
- name: outputs/llama3.3-miki-lora-out
  results: []
---

<!-- This model card has been generated automatically according to the information the Trainer had access to. You
should probably proofread and complete it, then remove this comment. -->

[<img src="https://raw.githubusercontent.com/axolotl-ai-cloud/axolotl/main/image/axolotl-badge-web.png" alt="Built with Axolotl" width="200" height="32"/>](https://github.com/axolotl-ai-cloud/axolotl)
<details><summary>See axolotl config</summary>

axolotl version: `0.15.0.dev0`
```yaml
base_model: aeon37/Llama-3.3-8B-Instruct-128K-heretic
#deepspeed: zero1_torch_compile.json # deepspeed_configs/zero1.json

# Automatically upload checkpoint and final model to HF
# hub_model_id: username/custom_model_name

model_type: LlamaForCausalLM
tokenizer_type: AutoTokenizer

trust_remote_code: true

  #plugins:
  # - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin

load_in_8bit: false
load_in_4bit: true 


datasets:
  - path: ramendik/kimify-ifeval-like
    type: chat_template
    # drop_system_message: true
    field_messages: messages
    # roles_to_train: ["assistant", "user"]
  - path: ramendik/kimify-20251115
    type: chat_template
    # drop_system_message: true
    field_messages: messages
  - path: ramendik/kimify-short-20260131
    type: chat_template
    # drop_system_message: true
    field_messages: messages


dataset_prepared_path: last_run_prepared_kimi
val_set_size: 0.1
output_dir: ./outputs/llama3.3-miki-lora-out
special_tokens:
 pad_token: <|end_of_text|>

adapter: qlora
lora_model_dir:

sequence_len: 2048
sample_packing: true

lora_r: 16
lora_alpha: 8
lora_dropout: 0.05
lora_target_linear: true
lora_modules_to_save:
  - embed_tokens
  - lm_head

use_wandb: true
wandb_project: Llama-3.3-8B-Instruct-128k-heretic-Kimi-miki
wandb_entity: 
wandb_watch:
wandb_name:
wandb_log_model:

gradient_accumulation_steps: 4
micro_batch_size: 2
num_epochs: 3
optimizer: paged_adamw_8bit
lr_scheduler: cosine
learning_rate: 0.0002

bf16: auto
tf32: false

gradient_checkpointing: true
gradient_checkpointing_kwargs: 
    use_reentrant: false
resume_from_checkpoint:
logging_steps: 1
flash_attention: true

warmup_ratio: 0.1
evals_per_epoch: 0
saves_per_epoch: 5 

save_first_step: true  # uncomment this to validate checkpoint saving works with your config

```

</details><br>

# outputs/llama3.3-miki-lora-out

This model is a fine-tuned version of [aeon37/Llama-3.3-8B-Instruct-128K-heretic](https://huggingface.co/aeon37/Llama-3.3-8B-Instruct-128K-heretic) on the ramendik/kimify-ifeval-like, the ramendik/kimify-20251115 and the ramendik/kimify-short-20260131 datasets.

## Model description

Without system prompts. Provided Lora so you can check what strenght to merge it with the model for best results. 
Untested !!

## Intended uses & limitations

More information needed

`llama-server -hf noctrex/Llama-3.3-8B-Instruct-128k-abliterated-GGUF:Q8_0  --lora-scaled rekrek/Llama-3.3-8B-Instruct-128K-heretic-kimified-lora 0.5 --lora-scaled rekrek/Llama-3.3-8B-Instruct-128K-heretic-kimified-lora 0.7 --lora-scaled rekrek/Llama-3.3-8B-Instruct-128K-heretic-kimified-lora 1.0 --port 8111 --host 0.0.0.0 -n 128000`

## Training and evaluation data

More information needed

`paged_adamw_8bit` can raise the batch size a bit more for quicker training, but results in some spikes.

## Training procedure

### Training hyperparameters

The following hyperparameters were used during training:
- learning_rate: 0.0002
- train_batch_size: 2
- eval_batch_size: 2
- seed: 42
- distributed_type: multi-GPU
- num_devices: 2
- gradient_accumulation_steps: 4
- total_train_batch_size: 16
- total_eval_batch_size: 4
- optimizer: Use OptimizerNames.PAGED_ADAMW_8BIT with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
- lr_scheduler_type: cosine
- lr_scheduler_warmup_steps: 103
- training_steps: 1032

### Training results

<img src="Train_Loss.png" alt="Train/Loss" width="600" height="200"/>

Could have trained for 2.4 epoch, seems a bit overfill.

### Framework versions

- PEFT 0.18.1
- Transformers 5.0.0
- Pytorch 2.9.1+cu128
- Datasets 4.5.0
- Tokenizers 0.22.2