---
library_name: peft
license: llama3.2
base_model: meta-llama/Llama-3.2-3B-Instruct
tags:
- axolotl
- base_model:adapter:meta-llama/Llama-3.2-3B-Instruct
- lora
- transformers
datasets:
- psychopenguin/indian_legal_dataset_qna
pipeline_tag: text-generation
model-index:
- name: legal_llama3.2-3b-instruct
  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: `0.16.0.dev0`
```yaml
seed: 9
gradient_checkpointing: True
auto_resume_from_checkpoints: True
tokenizer_save_jinja_files: True
trust_remote_code: True
tokenizer_use_fast: True
load_best_model_at_end: true

base_model: meta-llama/Llama-3.2-3B-Instruct
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer
chat_template: tokenizer_default

datasets: 
  - path: psychopenguin/indian_legal_dataset_qna
    type: alpaca
    split: train
    

sample_packing: True               
pad_to_sequence_len: True 
sequence_len: 2048
bf16: True
tf32: True
adapter: lora
lora_target_linear: True
lora_r: 16
lora_alpha: 32
lora_dropout: 0.05

flash_attention: True
flash_attn_cross_entropy: True


optimizer: paged_adamw_8bit
learning_rate: 0.0002
lr_scheduler: cosine
gradient_accumulation_steps: 1
micro_batch_size: 4 # change according to gpu vram
num_epochs: 5    # mostly will not reach 5 
neftune_noise_alpha: 5
early_stopping_patience: 6
save_steps: 100
save_total_limit: 3

val_set_size: 0.05
eval_strategy: steps
eval_steps: 100

use_wandb: True
wandb_project: major_project
wandb_name: legal-lm

output_dir: ./final_model
merge_lora: True
hf_use_auth_token: True
hub_model_id: psychopenguin/legal_llama3.2-3b-instruct


```

</details><br>

# legal_llama3.2-3b-instruct

This model is a fine-tuned version of [meta-llama/Llama-3.2-3B-Instruct](https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct) on the psychopenguin/indian_legal_dataset_qna dataset.
It achieves the following results on the evaluation set:
- Loss: 1.1103
- Ppl: 3.0354
- Memory/max Active (gib): 12.08
- Memory/max Allocated (gib): 12.08
- Memory/device Reserved (gib): 19.03

## Model description

More information needed

## Intended uses & limitations

More information needed

## Training and evaluation data

More information needed

## Training procedure

### Training hyperparameters

The following hyperparameters were used during training:
- learning_rate: 0.0002
- train_batch_size: 4
- eval_batch_size: 4
- seed: 9
- 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: 66
- training_steps: 2215

### Training results

| Training Loss | Epoch  | Step | Validation Loss | Ppl    | Active (gib) | Allocated (gib) | Reserved (gib) |
|:-------------:|:------:|:----:|:---------------:|:------:|:------------:|:---------------:|:--------------:|
| No log        | 0      | 0    | 1.8405          | 6.2998 | 12.05        | 12.05           | 13.32          |
| 1.3598        | 0.2257 | 100  | 1.3577          | 3.8873 | 12.08        | 12.08           | 19.16          |
| 1.2943        | 0.4515 | 200  | 1.2594          | 3.5233 | 12.08        | 12.08           | 19.03          |
| 1.1917        | 0.6772 | 300  | 1.1918          | 3.2931 | 12.08        | 12.08           | 19.03          |
| 1.1383        | 0.9029 | 400  | 1.1406          | 3.1288 | 12.08        | 12.08           | 19.03          |
| 0.9999        | 1.1287 | 500  | 1.1115          | 3.0389 | 12.08        | 12.08           | 19.03          |
| 0.9648        | 1.3544 | 600  | 1.0786          | 2.9405 | 12.08        | 12.08           | 19.03          |
| 0.9499        | 1.5801 | 700  | 1.0521          | 2.8637 | 12.08        | 12.08           | 19.03          |
| 0.9443        | 1.8059 | 800  | 1.0255          | 2.7884 | 12.08        | 12.08           | 19.03          |
| 0.7065        | 2.0316 | 900  | 1.0495          | 2.8563 | 12.08        | 12.08           | 19.03          |
| 0.7210        | 2.2573 | 1000 | 1.0306          | 2.8026 | 12.08        | 12.08           | 19.03          |
| 0.7406        | 2.4831 | 1100 | 1.0208          | 2.7753 | 12.08        | 12.08           | 19.03          |
| 0.6923        | 2.7088 | 1200 | 1.0075          | 2.7386 | 12.08        | 12.08           | 19.03          |
| 0.6635        | 2.9345 | 1300 | 0.9931          | 2.6997 | 12.08        | 12.08           | 19.03          |
| 0.5177        | 3.1603 | 1400 | 1.0536          | 2.8679 | 12.08        | 12.08           | 19.03          |
| 0.5290        | 3.3860 | 1500 | 1.0481          | 2.8524 | 12.08        | 12.08           | 19.03          |
| 0.5298        | 3.6117 | 1600 | 1.0446          | 2.8424 | 12.08        | 12.08           | 19.03          |
| 0.4862        | 3.8375 | 1700 | 1.0451          | 2.8438 | 12.08        | 12.08           | 19.03          |
| 0.3959        | 4.0632 | 1800 | 1.1158          | 3.0520 | 12.08        | 12.08           | 19.03          |
| 0.4204        | 4.2889 | 1900 | 1.1103          | 3.0354 | 12.08        | 12.08           | 19.03          |


### Framework versions

- PEFT 0.18.1
- Transformers 5.3.0
- Pytorch 2.9.1+cu128
- Datasets 4.5.0
- Tokenizers 0.22.2