NOTE: Read the Control Adapter documentation for implementation details.
Trained via qlora-pipe-lite:
# ==============================
# MODEL AND OUTPUT CONFIGURATION
# ==============================
model_dir = '/mnt/models/command-a-03-2025-uncut'
output_dir = '/mnt/fiction_finetunes/finetuned'
# ===========================
# TRAINING TYPE CONFIGURATION
# ===========================
use_control_adapters = true
load_in_4bit = true
# =============================
# CONTROL ADAPTER CONFIGURATION
# =============================
# ~20 tokens per trainable parameter (1e9/(64*64*(12288+1)))
lora_rank = 64
control_adapter_gamma = 0.1
# =======================
# OPTIMIZER CONFIGURATION
# =======================
lr = 1e-3
# ======================
# TRAINING CONFIGURATION
# ======================
sequence_len = 4096
pipeline_stages = 2
# 120 batch size (3*40) --> 480k tokens per step (4096*120)
gradient_accumulation_steps = 40
use_column_major_topology = true
# =====================
# DATASET CONFIGURATION
# =====================
sequence_prefix = 5 # "<BOS_TOKEN>"
document_prefix = [255000, 255007] # "<|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>"
document_suffix = 255001 # "<|END_OF_TURN_TOKEN|>"
mask_tokens = true # Mask all special tokens
drop_tails = true
mix_datasets = true
# -------------------
# POSITIVE CLASS DATA
# -------------------
[[datasets]]
dataset_path = '/mnt/datasets/books-fiction-paragraphs/*.json'
control_class = 1
# -------------------
# NEGATIVE CLASS DATA
# -------------------
[[datasets]]
dataset_path = '/mnt/datasets/slop-fiction-paragraphs/*.json'
control_class = -1
using ~1B tokens (ie: ~500M positive and ~500M negative) from:
- jukofyork/books-fiction-paragraphs (NOTE: I have not uploaded this, but you can use my jukofyork/gutenberg-fiction-paragraphs to get similar results...)
- jukofyork/slop-fiction-paragraphs
taking around 15 days using 6x RTX A6000 over 3 machines:
(hence the 120 batch size: (num_gpus / pipeline_stages) * gradient_accumulation_steps = (6 / 2) * 40 = 120)
NOTE: Ignore the 22.69 days - it crashed whilst I was away from home and had to be restarted a week later...
The control adapter was then converted to a LoRA using control_adapter_to_lora.py:
command-a-03-2025-writer-v1-lora
and then merged using the merge-lora space:
✓ Successfully merged and uploaded model!
Model URL: https://huggingface.co/jukofyork/command-a-03-2025-writer-v1
Scale factor: 1
Processed 49 shards
Merged 128 layers with LoRA weights
See command-a-03-2025-writer-v1-lora-gguf for a LoRA in GGUF format that can be used with the --lora option on top of the base command-a-03-2025-uncut to get the same effect.