See axolotl config

axolotl version: 0.15.0.dev0

# This is higher rank + human data
# + new data (the silence stuff).

# Dataset configuration (placeholder - will be populated at runtime)
datasets:
  - path: "newsessions_revisedcutdown.jsonl"
    type: "chat_template"
    field_messages: "conversations"
    message_property_mappings:
      role: "from"
      content: "value"
  - path: "hidden_compressedgamescontext_reducedctx.jsonl"
    type: "chat_template"
    field_messages: "conversations"
    message_property_mappings:
      role: "from"
      content: "value"
  - path: "manual_gameplay_truncated.jsonl"
    type: "chat_template"
    field_messages: "conversations"
    message_property_mappings:
      role: "from"
      content: "value"
  - path: "newdataplussilent-filtered.jsonl"
    type: "chat_template"
    field_messages: "conversations"
    message_property_mappings:
      role: "from"
      content: "value"
  - path: "newprompt_lessreduced_1.jsonl"
    type: "chat_template"
    field_messages: "conversations"
    message_property_mappings:
      role: "from"
      content: "value"
  - path: "newprompt_lessreduced_2.jsonl"
    type: "chat_template"
    field_messages: "conversations"
    message_property_mappings:
      role: "from"
      content: "value"
  - path: "newprompt_lessreduced_orig.jsonl"
    type: "chat_template"
    field_messages: "conversations"
    message_property_mappings:
      role: "from"
      content: "value"
  - path: "generics/bluemoon-2_5mil.jsonl"
    type: "chat_template"
    field_messages: "conversations"
    message_property_mappings:
      role: "from"
      content: "value"
  - path: "generics/capybara_2_5mil.jsonl"
    type: "chat_template"
    field_messages: "conversations"
    message_property_mappings:
      role: "from"
      content: "value"
  - path: "generics/qwq_3million.jsonl"
    type: "chat_template"
    field_messages: "conversations"
    message_property_mappings:
      role: "from"
      content: "value"

# Model configuration
base_model: "alpindale/Mistral-7B-v0.2-hf" # TODO EITHER MISTRAL OR QWEN BASE
tokenizer_type: "AutoTokenizer"
model_type: "AutoModelForCausalLM"
load_in_8bit: false
load_in_4bit: false
strict: false

# LoRA configuration
adapter: "lora"
lora_r: 128  # reducing this had helped
lora_alpha: 512  # maybe increase?
lora_dropout: 0.4  # increasing this had helped
lora_target_linear: true
lora_target_modules:
  - "gate_proj"
  - "down_proj"
  - "up_proj"
  - "q_proj"
  - "v_proj"
  - "k_proj"
  - "o_proj"

# Dataset preparation and output
dataset_prepared_path: "last_finetune_prepared"
output_dir: "./finetune-model-output"

# Training parameters
seed: 1337
sequence_len: 11000
sample_packing: false
pad_to_sequence_len: true
shuffle_merged_datasets: true

# Batch and gradient settings
gradient_accumulation_steps: 20
micro_batch_size: 6
eval_batch_size: 1

# Training duration
num_epochs: 2

# Optimizer and scheduler
optimizer: "paged_adamw_8bit"
lr_scheduler: "constant"
learning_rate: 0.0001  # as high as it can go
weight_decay: 0.01
max_grad_norm: 1  # TODO look into this further

# Training options
train_on_inputs: false
group_by_length: false
noisy_embedding_alpha: 0  # remove?

# Precision settings
bf16: true
fp16: false
tf32: false

# Memory optimization
gradient_checkpointing: true

# Attention mechanism
xformers_attention: false
flash_attention: true

# Chat template
chat_template: "chatml"

# Checkpointing and saving
auto_resume_from_checkpoints: false
save_strategy: "epoch"

# Evaluation
warmup_ratio: 0.1
evals_per_epoch: 1
val_set_size: 0.0
eval_sample_packing: false

# Logging
logging_steps: 1

# Weights & Biases (placeholders - will be populated if wandb_project is provided)
wandb_project: diplonations
wandb_entity: ""
wandb_watch: ""
wandb_run_id: ""
wandb_log_model: ""

# Special tokens
special_tokens:
  pad_token: "<unk>"
  eos_token: "</s>"

# Liger kernel optimizations
use_liger_kernel: true
plugins:
  - "axolotl.integrations.liger.LigerPlugin"
liger_rope: true
liger_rms_norm: true
liger_glu_activation: true
liger_layer_norm: true
liger_fused_linear_cross_entropy: true

hub_model_id: Heralax/dpn-newdata-normal-actual
hub_strategy: all_checkpoints

dpn-newdata-normal-actual

This model is a fine-tuned version of alpindale/Mistral-7B-v0.2-hf on the newsessions_revisedcutdown.jsonl, the hidden_compressedgamescontext_reducedctx.jsonl, the manual_gameplay_truncated.jsonl, the newdataplussilent-filtered.jsonl, the newprompt_lessreduced_1.jsonl, the newprompt_lessreduced_2.jsonl, the newprompt_lessreduced_orig.jsonl, the generics/bluemoon-2_5mil.jsonl, the generics/capybara_2_5mil.jsonl and the generics/qwq_3million.jsonl datasets.

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: 0.0001
• train_batch_size: 6
• eval_batch_size: 1
• seed: 1337
• gradient_accumulation_steps: 20
• total_train_batch_size: 120
• 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: constant
• lr_scheduler_warmup_steps: 65
• training_steps: 656

Training results

Framework versions

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