File size: 6,280 Bytes
c8f983f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
---
viewer: false
tags:
  - uv-script
  - unsloth
  - training
  - hf-jobs
  - vlm
  - fine-tuning
---

# 🦥 Unsloth Training Scripts for HF Jobs

UV scripts for fine-tuning LLMs and VLMs using [Unsloth](https://github.com/unslothai/unsloth) on [HF Jobs](https://huggingface.co/docs/hub/jobs) (on-demand cloud GPUs). UV handles dependency installation automatically, so you can run these scripts directly without any local setup.

These scripts can also be used or adapted by agents to train models for you.

## Prerequisites

- A Hugging Face account
- The [HF CLI](https://huggingface.co/docs/huggingface_hub/main/en/guides/cli) installed and authenticated (`hf auth login`)
- A dataset on the Hub in the appropriate format (see format requirements below). A strong LLM agent can often convert your data into the right format if needed.

## Data Formats

### LLM Fine-tuning (SFT)

Requires conversation data in ShareGPT or similar format:

```python
{
    "messages": [
        {"from": "human", "value": "What is the capital of France?"},
        {"from": "gpt", "value": "The capital of France is Paris."}
    ]
}
```

The script auto-converts common formats (ShareGPT, Alpaca, etc.) via `standardize_data_formats`. See [mlabonne/FineTome-100k](https://huggingface.co/datasets/mlabonne/FineTome-100k) for a working dataset example.

### VLM Fine-tuning

Requires `images` and `messages` columns:

```python
{
    "images": [<PIL.Image>],  # List of images
    "messages": [
        {
            "role": "user",
            "content": [
                {"type": "image"},
                {"type": "text", "text": "What's in this image?"}
            ]
        },
        {
            "role": "assistant",
            "content": [
                {"type": "text", "text": "A golden retriever playing fetch in a park."}
            ]
        }
    ]
}
```

See [davanstrien/iconclass-vlm-sft](https://huggingface.co/datasets/davanstrien/iconclass-vlm-sft) for a working dataset example, and [davanstrien/iconclass-vlm-qwen3-best](https://huggingface.co/davanstrien/iconclass-vlm-qwen3-best) for a model trained with these scripts.

### Continued Pretraining

Any dataset with a text column:

```python
{"text": "Your domain-specific text here..."}
```

Use `--text-column` if your column has a different name.

## Usage

View available options for any script:

```bash
uv run https://huggingface.co/datasets/unsloth/jobs/raw/main/sft-lfm2.5.py --help
```

### LLM fine-tuning

Fine-tune [LFM2.5-1.2B-Instruct](https://huggingface.co/LiquidAI/LFM2.5-1.2B-Instruct), a compact and efficient text model from Liquid AI:

```bash
hf jobs uv run \
  https://huggingface.co/datasets/unsloth/jobs/raw/main/sft-lfm2.5.py \
  --flavor a10g-small --secrets HF_TOKEN --timeout 4h \
  -- --dataset mlabonne/FineTome-100k \
     --num-epochs 1 \
     --eval-split 0.2 \
     --output-repo your-username/lfm-finetuned
```

### VLM fine-tuning

```bash
hf jobs uv run \
  https://huggingface.co/datasets/unsloth/jobs/raw/main/sft-qwen3-vl.py \
  --flavor a100-large --secrets HF_TOKEN \
  -- --dataset your-username/dataset \
     --trackio-space your-username/trackio \
     --output-repo your-username/my-model
```

### Continued pretraining

```bash
hf jobs uv run \
  https://huggingface.co/datasets/unsloth/jobs/raw/main/continued-pretraining.py \
  --flavor a100-large --secrets HF_TOKEN \
  -- --dataset your-username/domain-corpus \
     --text-column content \
     --max-steps 1000 \
     --output-repo your-username/domain-llm
```

### With Trackio monitoring

```bash
hf jobs uv run \
  https://huggingface.co/datasets/unsloth/jobs/raw/main/sft-lfm2.5.py \
  --flavor a10g-small --secrets HF_TOKEN \
  -- --dataset mlabonne/FineTome-100k \
     --trackio-space your-username/trackio \
     --output-repo your-username/lfm-finetuned
```

## Scripts

| Script                                                 | Base Model           | Task                          |
| ------------------------------------------------------ | -------------------- | ----------------------------- |
| [`sft-lfm2.5.py`](sft-lfm2.5.py)                       | LFM2.5-1.2B-Instruct | LLM fine-tuning (recommended) |
| [`sft-qwen3-vl.py`](sft-qwen3-vl.py)                   | Qwen3-VL-8B          | VLM fine-tuning               |
| [`sft-gemma3-vlm.py`](sft-gemma3-vlm.py)               | Gemma 3 4B           | VLM fine-tuning (smaller)     |
| [`continued-pretraining.py`](continued-pretraining.py) | Qwen3-0.6B           | Domain adaptation             |

## Common Options

| Option                    | Description                            | Default      |
| ------------------------- | -------------------------------------- | ------------ |
| `--dataset`               | HF dataset ID                          | _required_   |
| `--output-repo`           | Where to save trained model            | _required_   |
| `--max-steps`             | Number of training steps               | 500          |
| `--num-epochs`            | Train for N epochs instead of steps    | -            |
| `--eval-split`            | Fraction for evaluation (e.g., 0.2)    | 0 (disabled) |
| `--batch-size`            | Per-device batch size                  | 2            |
| `--gradient-accumulation` | Gradient accumulation steps            | 4            |
| `--lora-r`                | LoRA rank                              | 16           |
| `--learning-rate`         | Learning rate                          | 2e-4         |
| `--merge-model`           | Upload merged model (not just adapter) | false        |
| `--trackio-space`         | HF Space for live monitoring           | -            |
| `--run-name`              | Custom name for Trackio run            | auto         |

## Tips

- Use `--max-steps 10` to verify everything works before a full run
- `--eval-split 0.1` helps detect overfitting
- Run `hf jobs hardware` to see GPU pricing (A100-large ~$2.50/hr, L40S ~$1.80/hr)
- Add `--streaming` for very large datasets
- First training step may take a few minutes (CUDA kernel compilation)

## Links

- [HF Jobs Quickstart](https://huggingface.co/docs/hub/jobs-quickstart)
- [Unsloth Documentation](https://docs.unsloth.ai/)
- [UV Scripts Guide](https://docs.astral.sh/uv/guides/scripts/)