---
inference: false
library_name: transformers
language:
- en
- fr
- de
- es
- it
- pt
- ja
- ko
- zh
- ar
- el
- fa
- pl
- id
- cs
- he
- hi
- nl
- ro
- ru
- tr
- uk
- vi
license: cc-by-nc-4.0
base_model:
- CohereLabs/c4ai-command-a-03-2025
datasets:
- jukofyork/instruction-refusals-500MB
- jukofyork/instruction-responses-500MB
---

**NOTE**: Read the [Control Adapter documentation](https://github.com/jukofyork/qlora-pipe-lite/blob/main/docs/ControlAdapters.md) for implementation details.

---

Trained via [qlora-pipe-lite](https://github.com/jukofyork/qlora-pipe-lite):

```toml
# ==============================
# MODEL AND OUTPUT CONFIGURATION
# ==============================

model_dir = '/mnt/models/command-a-03-2025'
output_dir = '/mnt/finetunes/command-a-03-2025-uncut'

# ===========================
# TRAINING TYPE CONFIGURATION
# ===========================

use_control_adapters = true

load_in_4bit = true

# =============================
# CONTROL ADAPTER CONFIGURATION
# =============================

# ~4 tokens per trainable parameter (2e8/(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

# 30 batch size (3*10) --> 120k tokens per step (4096*30)
gradient_accumulation_steps = 10

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/instruction-responses-500MB/*.json'
control_class = 1

# -------------------
# NEGATIVE CLASS DATA
# -------------------

[[datasets]]
dataset_path = '/mnt/datasets/instruction-refusals-500MB/*.json'
control_class = -1
```

using ~200M tokens (ie: ~100M positive and ~100M negative) from:

- [jukofyork/instruction-responses-500MB](https://huggingface.co/datasets/jukofyork/instruction-responses-500MB)
- [jukofyork/instruction-refusals-500MB](https://huggingface.co/datasets/jukofyork/instruction-refusals-500MB)

taking just under 4 days using 6x `RTX A6000` over 3 machines:

![image/png](https://cdn-uploads.huggingface.co/production/uploads/65995c45539c808e84c38bf1/bOMISzLsmjimXDZ2k72Mu.png)

![image/png](https://cdn-uploads.huggingface.co/production/uploads/65995c45539c808e84c38bf1/9f1YbcE2nuFSR80Xpgl2H.png)

![image/png](https://cdn-uploads.huggingface.co/production/uploads/65995c45539c808e84c38bf1/7BJJK6UKRDCgqEqEIwlkA.png)

(hence the 30 batch size: `(num_gpus / pipeline_stages) * gradient_accumulation_steps = (6 / 2) * 10 = 30`)

---

The control adapter was then converted to a LoRA using [control_adapter_to_lora.py](https://github.com/jukofyork/qlora-pipe-lite/blob/main/control_adapter_to_lora.py):

[jukofyork/command-a-03-2025-uncut-lora](https://huggingface.co/jukofyork/command-a-03-2025-uncut-lora)

and then merged using the [merge-lora](https://huggingface.co/spaces/jukofyork/merge-lora) space.

---

See [https://huggingface.co/jukofyork/command-a-03-2025-uncut-lora-gguf](https://huggingface.co/jukofyork/command-a-03-2025-uncut-lora-gguf) for a LoRA in GGUF format that can be used with the `--lora` option on top of the base [CohereLabs/c4ai-command-a-03-2025](https://huggingface.co/CohereLabs/c4ai-command-a-03-2025) to get the same effect.