I decided to mess around with training methods again, considering the re-emegence of methods like multi-step training. Some people began doing it again, and so, why not? Inspired by AshhLimaRP's methology but done it my way.

Freya-S1

LoRA Trained on ~1.1GB of literature and raw text over Qwen 2.5's base model.
Cleaned text and literature as best as I could, still, may have had issues here and there.

Freya-S2

The first LoRA was applied over Qwen 2.5 Instruct, then I trained on top of that.
Reduced LoRA rank because it's mainly instruct and other details I won't get into.

Recommended Model Settings | Look, I just use these, they work fine enough. I don't even know how DRY or other meme samplers work. Your system prompt matters more anyway.

Prompt Format: ChatML
Temperature: 1+ # I don't know, man.
min_p: 0.05

Training time in total was ~10 Hours on a 8xH100 Node, sponsored by the Government of Singapore or something. Thanks for the national service allowance, MHA.

https://sao10k.carrd.co/ for contact.

See axolotl config

axolotl version: 0.6.0

yaml

base_model:
- s1: Qwen/Qwen2.5-14B
- s2: Qwen/Qwen2.5-14B-Instruct
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer

load_in_8bit: false
load_in_4bit: false
strict: false
sequence_len: 16384
bf16: auto
fp16:
tf32: false
flash_attention: true
special_tokens:
  
adapter: lora # 16-bit
lora_r:
- s1: 64
- s2: 32
lora_alpha: 64
lora_dropout: 0.2
lora_fan_in_fan_out:
peft_use_rslora: true
lora_target_linear: true
  
# Data
dataset_prepared_path: dataset_run_freya
datasets:
# S1 - Writing / Completion
  - path: datasets/eBooks-cleaned-75K
    type: completion
  - path: datasets/novels-clean-dedupe-10K
    type: completion
# S2 - Instruct
  - path: datasets/10k-amoral-full-fixed-sys.json
    type: chat_template
    chat_template: chatml
    roles_to_train: ["gpt"]
    field_messages: conversations
    message_field_role: from
    message_field_content: value
    train_on_eos: turn
  - path: datasets/44k-hespera-smartshuffle.json
    type: chat_template
    chat_template: chatml
    roles_to_train: ["gpt"]
    field_messages: conversations
    message_field_role: from
    message_field_content: value
    train_on_eos: turn
  - path: datasets/5k_rpg_adventure_instruct-sys.json
    type: chat_template
    chat_template: chatml
    roles_to_train: ["gpt"]
    field_messages: conversations
    message_field_role: from
    message_field_content: value
    train_on_eos: turn
shuffle_merged_datasets: true
warmup_ratio: 0.1

plugins:
  - axolotl.integrations.liger.LigerPlugin
liger_rope: true
liger_rms_norm: true
liger_layer_norm: true
liger_glu_activation: true
liger_fused_linear_cross_entropy: true

# Iterations
num_epochs:
- s1: 1
- s2: 2

# Sampling
sample_packing: true
pad_to_sequence_len: true
train_on_inputs: false
group_by_length: false

# Batching
gradient_accumulation_steps: 4
micro_batch_size: 2
gradient_checkpointing: unsloth

# Evaluation
val_set_size: 0.025
evals_per_epoch: 5
eval_table_size:
eval_max_new_tokens: 256
eval_sample_packing: false
eval_batch_size: 1

# Optimizer
optimizer: paged_ademamix_8bit
lr_scheduler: cosine
learning_rate:
- s1: 0.000002
- s2: 0.000004
weight_decay: 0.2
max_grad_norm: 10.0

# Garbage Collection
gc_steps: 10

# Misc
deepspeed: ./deepspeed_configs/zero2.json

Sao10K/14B-Qwen2.5-Freya-x1

14B-Qwen2.5-Freya-v1

Sao10K/
14B-Qwen2.5-Freya-x1