magistrate-3.2-3b-base

Magistrate 3.2 3B

Continued pretraining applied to meta-llama/Llama-3.2-3B using no synthetic legal data. ~250M tokens.

The model achieves the following results on the evaluation set:

• Loss: 0.6802

Instruct version is available here

<img src="https://raw.githubusercontent.com/axolotl-ai-cloud/axolotl/main/image/axolotl-badge-web.png" alt="Built with Axolotl" width="200" height="32"/>

<details><summary>See axolotl config</summary>

axolotl version: 0.4.1

base_model: meta-llama/Llama-3.2-3B
model_type: LlamaForCausalLM
tokenizer_type: AutoTokenizer

load_in_8bit: false
load_in_4bit: false
strict: false

datasets:
  - path: json
    data_files: "data/amendments_with_content_converted.json"
    type: completion
  - path: json
    data_files: "data/federal_rules_converted.json"
    type: completion
  - path: json
    data_files: "data/cornell_legal_encyclopedias_converted.json"
    type: completion
  - path: json
    data_files: "data/pocket_guide_for_judges_converted.json"
    type: completion
  - path: json
    data_files: "data/us_federal_code.json"
    type: completion
  - path: json
    data_files: "data/us_supreme_court_summaries_converted.json"
    type: completion
  - path: json
    data_files: "data/us_supreme_court_converted.json"
    type: completion
  - path: json
    data_files: "data/ucfr.json"
    type: completion
  - path: json
    data_files: "data/map-code-filtered.json"
    type: completion
  
dataset_prepared_path:
val_set_size: 0.05
output_dir: ./outputs/lora-out

sequence_len: 8192
sample_packing: true
eval_sample_packing: false
pad_to_sequence_len: true

# adapter: lora
# lora_model_dir:
# lora_r: 128
# lora_alpha: 32
# lora_dropout: 0.05
# lora_target_linear: true
# lora_fan_in_fan_out:
# lora_modules_to_save:
#   - embed_tokens
#   - lm_head

unfrozen_parameters:
- ^lm_head.weight$
- ^model.embed_tokens.weight$
# mlp.down_proj layers
- model.layers.0.mlp.down_proj
- model.layers.1.mlp.down_proj
- model.layers.17.mlp.down_proj
- model.layers.19.mlp.down_proj
- model.layers.18.mlp.down_proj
- model.layers.5.mlp.down_proj
- model.layers.20.mlp.down_proj
- model.layers.2.mlp.down_proj
- model.layers.4.mlp.down_proj
- model.layers.6.mlp.down_proj
- model.layers.3.mlp.down_proj
- model.layers.16.mlp.down_proj
- model.layers.15.mlp.down_proj
- model.layers.13.mlp.down_proj
# mlp.gate_proj layers
- model.layers.0.mlp.gate_proj
- model.layers.1.mlp.gate_proj
- model.layers.2.mlp.gate_proj
- model.layers.3.mlp.gate_proj
- model.layers.22.mlp.gate_proj
- model.layers.21.mlp.gate_proj
- model.layers.20.mlp.gate_proj
- model.layers.23.mlp.gate_proj
- model.layers.19.mlp.gate_proj
- model.layers.4.mlp.gate_proj
- model.layers.18.mlp.gate_proj
- model.layers.17.mlp.gate_proj
- model.layers.5.mlp.gate_proj
- model.layers.24.mlp.gate_proj
# mlp.up_proj layers
- model.layers.4.mlp.up_proj
- model.layers.3.mlp.up_proj
- model.layers.5.mlp.up_proj
- model.layers.6.mlp.up_proj
- model.layers.7.mlp.up_proj
- model.layers.2.mlp.up_proj
- model.layers.8.mlp.up_proj
- model.layers.14.mlp.up_proj
- model.layers.13.mlp.up_proj
- model.layers.11.mlp.up_proj
- model.layers.9.mlp.up_proj
- model.layers.1.mlp.up_proj
- model.layers.15.mlp.up_proj
- model.layers.12.mlp.up_proj
# self_attn.k_proj layers
- model.layers.25.self_attn.k_proj
- model.layers.22.self_attn.k_proj
- model.layers.19.self_attn.k_proj
- model.layers.20.self_attn.k_proj
- model.layers.17.self_attn.k_proj
- model.layers.24.self_attn.k_proj
- model.layers.23.self_attn.k_proj
- model.layers.18.self_attn.k_proj
- model.layers.21.self_attn.k_proj
- model.layers.27.self_attn.k_proj
- model.layers.15.self_attn.k_proj
- model.layers.10.self_attn.k_proj
- model.layers.6.self_attn.k_proj
- model.layers.5.self_attn.k_proj
# self_attn.o_proj layers

wandb_project:
wandb_entity:
wandb_watch:
wandb_name:
wandb_log_model:

gradient_accumulation_steps: 4
micro_batch_size: 2
num_epochs: 3
optimizer: paged_adamw_32bit

# Gradient clipping max norm
max_grad_norm: 1.0
noisy_embedding_alpha: 0 # no noisy embedding to ensure maximal memorization 


lr_scheduler: cosine
learning_rate: 0.0002
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: 690
evals_per_epoch: 2
eval_table_size:
eval_max_new_tokens: 128
saves_per_epoch: 1
debug:
deepspeed: deepspeed_configs/zero3.json
weight_decay: 0.0
fsdp:
fsdp_config:
special_tokens:
  pad_token: <|end_of_text|>

</details><br>

Model description

This is a base model trained on US Supreme Court proceedings, US federal code and regulations.

Intended uses & limitations

This model is intended for research purposes. You are liable for all model outputs.

Training and evaluation data

The training data consists of US Supreme Court verdicts, federal regulations, laws and treaties.

Some other resources have been included from institutions like CLL to fill in the gaps in knowledge for industry jargon.

Training procedure

Spectrum top 35% fine tune. Thanks to the cognitive computations team for the work done on spectrum.

Methodology based on Cohere's paper: To Code, or Not To Code? Exploring Impact of Code in Pre-training

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: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 690
• num_epochs: 3

Training results

Framework versions

• Transformers 4.45.0
• Pytorch 2.3.1+cu121
• Datasets 2.21.0
• Tokenizers 0.20.0