Commit
Β·
39535cd
1
Parent(s):
b4c35f6
viewer false
Browse files
README.md
CHANGED
@@ -1,3 +1,7 @@
|
|
|
|
|
|
|
|
|
|
1 |
# π OpenAI GPT OSS Models - Simple Generation Script
|
2 |
|
3 |
Generate synthetic datasets using OpenAI's GPT OSS models with transparent reasoning. Works on HuggingFace Jobs with L4 GPUs!
|
@@ -21,26 +25,28 @@ hf jobs uv run --flavor l4x4 --secrets HF_TOKEN=hf_*** \
|
|
21 |
|
22 |
## π Script Options
|
23 |
|
24 |
-
| Option
|
25 |
-
|
26 |
-
| `--input-dataset`
|
27 |
-
| `--output-dataset`
|
28 |
-
| `--prompt-column`
|
29 |
-
| `--model-id`
|
30 |
-
| `--max-samples`
|
31 |
-
| `--max-new-tokens`
|
32 |
-
| `--reasoning-effort` | Reasoning depth: low/medium/high | `medium`
|
33 |
-
| `--temperature`
|
34 |
-
| `--top-p`
|
35 |
|
36 |
**Note**: `max-new-tokens` auto-scales based on `reasoning-effort` if not set:
|
|
|
37 |
- `low`: 512 tokens
|
38 |
-
- `medium`: 1024 tokens
|
39 |
- `high`: 2048 tokens (prevents truncation of detailed reasoning)
|
40 |
|
41 |
## π‘ What You Get
|
42 |
|
43 |
The output dataset contains:
|
|
|
44 |
- `prompt`: Original prompt from input dataset
|
45 |
- `raw_output`: Full model response with channel markers
|
46 |
- `model`: Model ID used
|
@@ -49,10 +55,12 @@ The output dataset contains:
|
|
49 |
### Understanding the Output
|
50 |
|
51 |
The raw output contains special channel markers:
|
|
|
52 |
- `<|channel|>analysis<|message|>` - Chain of thought reasoning
|
53 |
- `<|channel|>final<|message|>` - The actual response
|
54 |
|
55 |
Example raw output structure:
|
|
|
56 |
```
|
57 |
<|channel|>analysis<|message|>
|
58 |
[Reasoning about the task...]
|
@@ -65,6 +73,7 @@ Example raw output structure:
|
|
65 |
### Test with Different Reasoning Levels
|
66 |
|
67 |
**High reasoning (most detailed):**
|
|
|
68 |
```bash
|
69 |
hf jobs uv run --flavor l4x4 --secrets HF_TOKEN=hf_*** \
|
70 |
https://huggingface.co/datasets/davanstrien/openai-oss/raw/main/gpt_oss_minimal.py \
|
@@ -76,6 +85,7 @@ hf jobs uv run --flavor l4x4 --secrets HF_TOKEN=hf_*** \
|
|
76 |
```
|
77 |
|
78 |
**Low reasoning (fastest):**
|
|
|
79 |
```bash
|
80 |
hf jobs uv run --flavor l4x4 --secrets HF_TOKEN=hf_*** \
|
81 |
https://huggingface.co/datasets/davanstrien/openai-oss/raw/main/gpt_oss_minimal.py \
|
@@ -88,29 +98,34 @@ hf jobs uv run --flavor l4x4 --secrets HF_TOKEN=hf_*** \
|
|
88 |
|
89 |
## π₯οΈ GPU Requirements
|
90 |
|
91 |
-
| Model
|
92 |
-
|
93 |
-
| **openai/gpt-oss-20b**
|
94 |
-
| **openai/gpt-oss-120b** | ~240GB
|
95 |
|
96 |
**Note**: The 20B model automatically dequantizes from MXFP4 to bf16 on non-Hopper GPUs, requiring more memory than the quantized size.
|
97 |
|
98 |
## π§ Technical Details
|
99 |
|
100 |
### Why L4x4?
|
|
|
101 |
- The 20B model needs ~40GB VRAM when dequantized
|
102 |
- Single A10G (24GB) is insufficient
|
103 |
- L4x4 provides 96GB total memory across 4 GPUs
|
104 |
- Cost-effective compared to A100 instances
|
105 |
|
106 |
### Reasoning Effort
|
|
|
107 |
The `reasoning_effort` parameter controls how much chain-of-thought reasoning the model generates:
|
|
|
108 |
- `low`: Quick responses with minimal reasoning
|
109 |
- `medium`: Balanced reasoning (default)
|
110 |
- `high`: Detailed step-by-step reasoning
|
111 |
|
112 |
### Sampling Parameters
|
|
|
113 |
OpenAI recommends `temperature=1.0` and `top_p=1.0` as defaults for GPT OSS models:
|
|
|
114 |
- These settings provide good diversity without compromising quality
|
115 |
- The model was trained to work well with these parameters
|
116 |
- Adjust only if you need specific behavior (e.g., lower temperature for more deterministic output)
|
@@ -123,4 +138,4 @@ OpenAI recommends `temperature=1.0` and `top_p=1.0` as defaults for GPT OSS mode
|
|
123 |
|
124 |
---
|
125 |
|
126 |
-
|
|
|
1 |
+
---
|
2 |
+
viewer: false
|
3 |
+
---
|
4 |
+
|
5 |
# π OpenAI GPT OSS Models - Simple Generation Script
|
6 |
|
7 |
Generate synthetic datasets using OpenAI's GPT OSS models with transparent reasoning. Works on HuggingFace Jobs with L4 GPUs!
|
|
|
25 |
|
26 |
## π Script Options
|
27 |
|
28 |
+
| Option | Description | Default |
|
29 |
+
| -------------------- | -------------------------------- | -------------------------- |
|
30 |
+
| `--input-dataset` | HuggingFace dataset to process | Required |
|
31 |
+
| `--output-dataset` | Output dataset name | Required |
|
32 |
+
| `--prompt-column` | Column containing prompts | `prompt` |
|
33 |
+
| `--model-id` | Model to use | `openai/gpt-oss-20b` |
|
34 |
+
| `--max-samples` | Limit samples to process | None (all) |
|
35 |
+
| `--max-new-tokens` | Max tokens to generate | Auto-scales: 512/1024/2048 |
|
36 |
+
| `--reasoning-effort` | Reasoning depth: low/medium/high | `medium` |
|
37 |
+
| `--temperature` | Sampling temperature | `1.0` |
|
38 |
+
| `--top-p` | Top-p sampling | `1.0` |
|
39 |
|
40 |
**Note**: `max-new-tokens` auto-scales based on `reasoning-effort` if not set:
|
41 |
+
|
42 |
- `low`: 512 tokens
|
43 |
+
- `medium`: 1024 tokens
|
44 |
- `high`: 2048 tokens (prevents truncation of detailed reasoning)
|
45 |
|
46 |
## π‘ What You Get
|
47 |
|
48 |
The output dataset contains:
|
49 |
+
|
50 |
- `prompt`: Original prompt from input dataset
|
51 |
- `raw_output`: Full model response with channel markers
|
52 |
- `model`: Model ID used
|
|
|
55 |
### Understanding the Output
|
56 |
|
57 |
The raw output contains special channel markers:
|
58 |
+
|
59 |
- `<|channel|>analysis<|message|>` - Chain of thought reasoning
|
60 |
- `<|channel|>final<|message|>` - The actual response
|
61 |
|
62 |
Example raw output structure:
|
63 |
+
|
64 |
```
|
65 |
<|channel|>analysis<|message|>
|
66 |
[Reasoning about the task...]
|
|
|
73 |
### Test with Different Reasoning Levels
|
74 |
|
75 |
**High reasoning (most detailed):**
|
76 |
+
|
77 |
```bash
|
78 |
hf jobs uv run --flavor l4x4 --secrets HF_TOKEN=hf_*** \
|
79 |
https://huggingface.co/datasets/davanstrien/openai-oss/raw/main/gpt_oss_minimal.py \
|
|
|
85 |
```
|
86 |
|
87 |
**Low reasoning (fastest):**
|
88 |
+
|
89 |
```bash
|
90 |
hf jobs uv run --flavor l4x4 --secrets HF_TOKEN=hf_*** \
|
91 |
https://huggingface.co/datasets/davanstrien/openai-oss/raw/main/gpt_oss_minimal.py \
|
|
|
98 |
|
99 |
## π₯οΈ GPU Requirements
|
100 |
|
101 |
+
| Model | Memory Required | Recommended Flavor |
|
102 |
+
| ----------------------- | --------------- | ---------------------- |
|
103 |
+
| **openai/gpt-oss-20b** | ~40GB | `l4x4` (4x24GB = 96GB) |
|
104 |
+
| **openai/gpt-oss-120b** | ~240GB | `8xa100` (8x80GB) |
|
105 |
|
106 |
**Note**: The 20B model automatically dequantizes from MXFP4 to bf16 on non-Hopper GPUs, requiring more memory than the quantized size.
|
107 |
|
108 |
## π§ Technical Details
|
109 |
|
110 |
### Why L4x4?
|
111 |
+
|
112 |
- The 20B model needs ~40GB VRAM when dequantized
|
113 |
- Single A10G (24GB) is insufficient
|
114 |
- L4x4 provides 96GB total memory across 4 GPUs
|
115 |
- Cost-effective compared to A100 instances
|
116 |
|
117 |
### Reasoning Effort
|
118 |
+
|
119 |
The `reasoning_effort` parameter controls how much chain-of-thought reasoning the model generates:
|
120 |
+
|
121 |
- `low`: Quick responses with minimal reasoning
|
122 |
- `medium`: Balanced reasoning (default)
|
123 |
- `high`: Detailed step-by-step reasoning
|
124 |
|
125 |
### Sampling Parameters
|
126 |
+
|
127 |
OpenAI recommends `temperature=1.0` and `top_p=1.0` as defaults for GPT OSS models:
|
128 |
+
|
129 |
- These settings provide good diversity without compromising quality
|
130 |
- The model was trained to work well with these parameters
|
131 |
- Adjust only if you need specific behavior (e.g., lower temperature for more deterministic output)
|
|
|
138 |
|
139 |
---
|
140 |
|
141 |
+
_Last tested: 2025-01-06 on HF Jobs with l4x4 flavor_
|