File size: 4,371 Bytes
78e2e1c
bbe5ce0
78e2e1c
bbe5ce0
78e2e1c
bbe5ce0
78e2e1c
bbe5ce0
a4ee9cd
bbe5ce0
a4ee9cd
78e2e1c
 
 
bbe5ce0
78e2e1c
bbe5ce0
78e2e1c
 
bbe5ce0
 
78e2e1c
bbe5ce0
78e2e1c
 
 
 
 
 
 
db35e01
78e2e1c
db35e01
 
 
 
 
 
 
bbe5ce0
78e2e1c
a4ee9cd
78e2e1c
 
 
 
 
bbe5ce0
78e2e1c
bbe5ce0
78e2e1c
 
 
bbe5ce0
78e2e1c
 
 
 
 
 
 
bbe5ce0
a4ee9cd
bbe5ce0
78e2e1c
 
 
bbe5ce0
78e2e1c
 
bbe5ce0
78e2e1c
bbe5ce0
78e2e1c
 
bbe5ce0
 
78e2e1c
bbe5ce0
78e2e1c
 
 
 
bbe5ce0
78e2e1c
 
a4ee9cd
 
78e2e1c
bbe5ce0
78e2e1c
 
 
 
bbe5ce0
78e2e1c
a4ee9cd
 
 
78e2e1c
 
 
 
 
bbe5ce0
78e2e1c
 
 
 
 
bbe5ce0
db35e01
 
 
 
 
 
78e2e1c
bbe5ce0
a4ee9cd
 
78e2e1c
bbe5ce0
 
 
78e2e1c
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
# πŸš€ OpenAI GPT OSS Models - Simple Generation Script

Generate synthetic datasets using OpenAI's GPT OSS models with transparent reasoning. Works on HuggingFace Jobs with L4 GPUs!

## βœ… Tested & Working

Successfully tested on HF Jobs with `l4x4` flavor (4x L4 GPUs = 96GB total memory).

## 🌟 Quick Start

```bash
# Run on HF Jobs (tested and working)
hf jobs uv run --flavor l4x4 --secrets HF_TOKEN=hf_*** \
    https://huggingface.co/datasets/davanstrien/openai-oss/raw/main/gpt_oss_minimal.py \
    --input-dataset davanstrien/haiku_dpo \
    --output-dataset username/gpt-oss-haiku \
    --prompt-column question \
    --max-samples 2 \
    --reasoning-effort high
```

## πŸ“‹ Script Options

| Option | Description | Default |
|--------|-------------|---------|
| `--input-dataset` | HuggingFace dataset to process | Required |
| `--output-dataset` | Output dataset name | Required |
| `--prompt-column` | Column containing prompts | `prompt` |
| `--model-id` | Model to use | `openai/gpt-oss-20b` |
| `--max-samples` | Limit samples to process | None (all) |
| `--max-new-tokens` | Max tokens to generate | Auto-scales: 512/1024/2048 |
| `--reasoning-effort` | Reasoning depth: low/medium/high | `medium` |
| `--temperature` | Sampling temperature | `1.0` |
| `--top-p` | Top-p sampling | `1.0` |

**Note**: `max-new-tokens` auto-scales based on `reasoning-effort` if not set:
- `low`: 512 tokens
- `medium`: 1024 tokens  
- `high`: 2048 tokens (prevents truncation of detailed reasoning)

## πŸ’‘ What You Get

The output dataset contains:
- `prompt`: Original prompt from input dataset
- `raw_output`: Full model response with channel markers
- `model`: Model ID used
- `reasoning_effort`: The reasoning level used

### Understanding the Output

The raw output contains special channel markers:
- `<|channel|>analysis<|message|>` - Chain of thought reasoning
- `<|channel|>final<|message|>` - The actual response

Example raw output structure:
```
<|channel|>analysis<|message|>
[Reasoning about the task...]
<|channel|>final<|message|>
[Actual haiku or response]
```

## 🎯 Examples

### Test with Different Reasoning Levels

**High reasoning (most detailed):**
```bash
hf jobs uv run --flavor l4x4 --secrets HF_TOKEN=hf_*** \
    https://huggingface.co/datasets/davanstrien/openai-oss/raw/main/gpt_oss_minimal.py \
    --input-dataset davanstrien/haiku_dpo \
    --output-dataset username/haiku-high \
    --prompt-column question \
    --reasoning-effort high \
    --max-samples 5
```

**Low reasoning (fastest):**
```bash
hf jobs uv run --flavor l4x4 --secrets HF_TOKEN=hf_*** \
    https://huggingface.co/datasets/davanstrien/openai-oss/raw/main/gpt_oss_minimal.py \
    --input-dataset davanstrien/haiku_dpo \
    --output-dataset username/haiku-low \
    --prompt-column question \
    --reasoning-effort low \
    --max-samples 10
```

## πŸ–₯️ GPU Requirements

| Model | Memory Required | Recommended Flavor |
|-------|----------------|-------------------|
| **openai/gpt-oss-20b** | ~40GB | `l4x4` (4x24GB = 96GB) |
| **openai/gpt-oss-120b** | ~240GB | `8xa100` (8x80GB) |

**Note**: The 20B model automatically dequantizes from MXFP4 to bf16 on non-Hopper GPUs, requiring more memory than the quantized size.

## πŸ”§ Technical Details

### Why L4x4?
- The 20B model needs ~40GB VRAM when dequantized
- Single A10G (24GB) is insufficient
- L4x4 provides 96GB total memory across 4 GPUs
- Cost-effective compared to A100 instances

### Reasoning Effort
The `reasoning_effort` parameter controls how much chain-of-thought reasoning the model generates:
- `low`: Quick responses with minimal reasoning
- `medium`: Balanced reasoning (default)
- `high`: Detailed step-by-step reasoning

### Sampling Parameters
OpenAI recommends `temperature=1.0` and `top_p=1.0` as defaults for GPT OSS models:
- These settings provide good diversity without compromising quality
- The model was trained to work well with these parameters
- Adjust only if you need specific behavior (e.g., lower temperature for more deterministic output)

## πŸ“š Resources

- [Model: openai/gpt-oss-20b](https://huggingface.co/openai/gpt-oss-20b)
- [HF Jobs Documentation](https://huggingface.co/docs/hub/spaces-gpu-jobs)
- [Dataset: davanstrien/haiku_dpo](https://huggingface.co/datasets/davanstrien/haiku_dpo)

---

*Last tested: 2025-01-06 on HF Jobs with l4x4 flavor*