Creation process: Upscale > Pretrain > SFT > DPO
All training was qlora (including pretrain).
Pretrained on 177MB of data. Dataset consisteted mostly of Light Novels, NSFW stories, SFW stories and filled out with general corpus text from Huggingface FineWeb-2 dataset.
The model then went through SFT using a dataset of approx 3.6 million tokens, 700 RP conversations, 1000 creative writing / instruct samples and about 100 summaries. The bulk of this data has been made public.
Finally, DPO was used to make the model more consistent.
>
Mergekit Config
base_model: anthracite-core/Mistral-Small-3.2-24B-Instruct-2506-Text-Only
merge_method: passthrough
dtype: bfloat16
slices:
- sources:
- model: anthracite-core/Mistral-Small-3.2-24B-Instruct-2506-Text-Only
layer_range: [0, 29]
- sources:
- model: anthracite-core/Mistral-Small-3.2-24B-Instruct-2506-Text-Only
layer_range: [10, 39]
>
Axolotl configs
Not optimized for cost / performance efficiency, YMMV.
SFT 1*H100
# ====================
# MODEL CONFIGURATION
# ====================
base_model: ./Upscale_Mistral-PT/merged
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer
chat_template: mistral_v7_tekken
# ====================
# DATASET CONFIGURATION
# ====================
datasets:
- path: ./dataset.jsonl
type: chat_template
split: train
chat_template_strategy: tokenizer
field_messages: messages
message_property_mappings:
role: role
content: content
roles:
user: ["user"]
assistant: ["assistant"]
system: ["system"]
dataset_prepared_path:
train_on_inputs: false # Only train on assistant responses
# ====================
# QLORA CONFIGURATION
# ====================
adapter: qlora
load_in_4bit: true
lora_r: 128
lora_alpha: 128
lora_dropout: 0.1
lora_target_linear: true
# lora_modules_to_save: # Uncomment only if you added NEW tokens
# ====================
# TRAINING PARAMETERS
# ====================
num_epochs: 2
micro_batch_size: 4
gradient_accumulation_steps: 2
learning_rate: 1.5e-5
optimizer: paged_adamw_8bit
lr_scheduler: rex
warmup_ratio: 0.05
weight_decay: 0.01
max_grad_norm: 1.0
# ====================
# SEQUENCE & PACKING
# ====================
sequence_len: 8192
sample_packing: true
eval_sample_packing: false
pad_to_sequence_len: true
# ====================
# HARDWARE OPTIMIZATIONS
# ====================
bf16: auto
flash_attention: true
gradient_checkpointing: true
# ====================
# EVALUATION & CHECKPOINTING
# ====================
save_strategy: steps
save_steps: 5
save_total_limit: 5 # Keep best + last few checkpoints
load_best_model_at_end: true
greater_is_better: false
# ====================
# LOGGING & OUTPUT
# ====================
output_dir: ./Upscale_Mistral-PT-SFT-2
logging_steps: 2
save_safetensors: true
# ====================
# WANDB TRACKING
# ====================
wandb_project: MS3-2-SFT
wandb_entity: your_entity
wandb_name: run_name
