Aurore-Reveil/seed-2.0-reasoning-reverse-ckpts

See axolotl config

axolotl version: 0.14.0

# === Model Configuration ===
base_model: PocketDoc/Dans-PersonalityEngine-V1.1.0-12b
load_in_8bit: false
load_in_4bit: false

# === Training Setup ===
num_epochs: 3
micro_batch_size: 2
gradient_accumulation_steps: 1
sequence_len: 2048
sample_packing: true
pad_to_sequence_len: true

# === Hyperparameter Configuration ===
optimizer: adamw_torch_8bit
learning_rate: 1e-5
lr_scheduler: cosine
weight_decay: 0.001
max_grad_norm: 0.1
warmup_ratio: 0.05
cosine_min_lr_ratio: 0.1

# === Data Configuration ===
datasets:
  - path: allura-forge/reasoning-trace-generator-dataset
    type: chat_template
    split: train
chat_template: tokenizer_default

dataset_prepared_path: last_run_prepared

# === Hardware Optimization ===
gradient_checkpointing: offload

# === Checkpointing ===
saves_per_epoch: 1

# === Advanced Settings ===
output_dir: ./model-output
bf16: auto
flash_attention: true
train_on_inputs: false
group_by_length: false
logging_steps: 1
trust_remote_code: false

plugins:
  - axolotl.integrations.liger.LigerPlugin
  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin

fsdp:
  - auto_wrap
  - full_shard
fsdp_config:
  fsdp_version: 2
  fsdp_offload_params: false
  fsdp_cpu_ram_efficient_loading: true
  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
  fsdp_transformer_layer_cls_to_wrap: MistralDecoderLayer
  fsdp_state_dict_type: SHARDED_STATE_DICT
  fsdp_sharding_strategy: FULL_SHARD
  fsdp_reshard_after_forward: true
  fsdp_activation_checkpointing: true # will disable if doesnt work

model-output

This model is a fine-tuned version of PocketDoc/Dans-PersonalityEngine-V1.1.0-12b on the allura-forge/reasoning-trace-generator-dataset 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: 1e-05
• train_batch_size: 2
• eval_batch_size: 2
• seed: 42
• distributed_type: multi-GPU
• num_devices: 8
• total_train_batch_size: 16
• total_eval_batch_size: 16
• optimizer: Use OptimizerNames.ADAMW_TORCH_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: 9
• training_steps: 186

Training results

Framework versions

• Transformers 5.0.0
• Pytorch 2.9.0+cu130
• Datasets 4.5.0
• Tokenizers 0.22.2