CrucibleLab-TG/M3.2-24B-loki-V2-chpts

See axolotl config

axolotl version: 0.13.0.dev0

base_model: anthracite-core/Mistral-Small-3.2-24B-Instruct-2506-Text-Only
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer
chat_template: tokenizer_default
trust_remote_code: true

special_tokens:
  pad_token: "<pad>"
  eos_token: "</s>"

datasets:
  - path: CrucibleLab/Loki_V2_Cleaned
    ds_type: json
    type: chat_template
    chat_template_strategy: tokenizer_default
    field_messages: conversations
    message_property_mappings:
      role: from
      content: value      
    roles:
      user: ["user"]
      assistant: ["assistant"]
      system: ["system"]
    roles_to_train: ["assistant"]

dataset_prepared_path: "last_run_prepared"
output_dir: /workspace/data/24b-Qlora
train_on_inputs: false
shuffle_merged_datasets: true

adapter: qlora
load_in_4bit: true
lora_r: 256
lora_alpha: 256
lora_target_linear: true
peft_use_rslora: true

sequence_len: 8192
sample_packing: true
pad_to_sequence_len: true

micro_batch_size: 8
gradient_accumulation_steps: 1
num_epochs: 1

lr_scheduler: rex
learning_rate: 2e-5
max_grad_norm: 4.5

save_steps: 1000        # Every 1000 steps
save_total_limit: 10    # Keep last 10 checkpoints
warmup_ratio: 0.05
hub_model_id: CrucibleLab/M3.2-24B-loki-V2
hub_strategy: all_checkpoints
eval_strategy: "no"

bf16: auto
tf32: true

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

optimizer: adamw_8bit

weight_decay: 0.0
save_safetensors: true

wandb_project: M3.2-24B-loki-V2
wandb_entity: CrucibleLabs
wandb_name: M3.2-24B-loki-V2

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

M3.2-24B-loki-V2

This model is a fine-tuned version of anthracite-core/Mistral-Small-3.2-24B-Instruct-2506-Text-Only on the CrucibleLab/Loki_V2_Cleaned dataset.

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: 2e-05
• train_batch_size: 8
• eval_batch_size: 8
• seed: 42
• optimizer: Use OptimizerNames.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: 389
• training_steps: 7797

Training results

Framework versions

• PEFT 0.18.0
• Transformers 4.57.1
• Pytorch 2.8.0+cu128
• Datasets 4.4.1
• Tokenizers 0.22.1