davanstrien HF Staff commited on
Commit
db35e01
·
1 Parent(s): 78e2e1c

Add temperature, top_p, and smart max_tokens defaults

Browse files

- Set temperature=1.0 and top_p=1.0 as defaults (OpenAI recommended)
- Add --temperature and --top-p flags for customization
- Auto-scale max_new_tokens based on reasoning_effort:
- low: 512 tokens
- medium: 1024 tokens
- high: 2048 tokens (prevents truncation)
- Document OpenAI's sampling recommendations in README
- Users can still override with explicit --max-new-tokens

Files changed (2) hide show
  1. README.md +14 -1
  2. gpt_oss_minimal.py +33 -10
README.md CHANGED
@@ -28,8 +28,15 @@ hf jobs uv run --flavor l4x4 --secrets HF_TOKEN=hf_*** \
28
  | `--prompt-column` | Column containing prompts | `prompt` |
29
  | `--model-id` | Model to use | `openai/gpt-oss-20b` |
30
  | `--max-samples` | Limit samples to process | None (all) |
31
- | `--max-new-tokens` | Max tokens to generate | `1024` |
32
  | `--reasoning-effort` | Reasoning depth: low/medium/high | `medium` |
 
 
 
 
 
 
 
33
 
34
  ## 💡 What You Get
35
 
@@ -102,6 +109,12 @@ The `reasoning_effort` parameter controls how much chain-of-thought reasoning th
102
  - `medium`: Balanced reasoning (default)
103
  - `high`: Detailed step-by-step reasoning
104
 
 
 
 
 
 
 
105
  ## 📚 Resources
106
 
107
  - [Model: openai/gpt-oss-20b](https://huggingface.co/openai/gpt-oss-20b)
 
28
  | `--prompt-column` | Column containing prompts | `prompt` |
29
  | `--model-id` | Model to use | `openai/gpt-oss-20b` |
30
  | `--max-samples` | Limit samples to process | None (all) |
31
+ | `--max-new-tokens` | Max tokens to generate | Auto-scales: 512/1024/2048 |
32
  | `--reasoning-effort` | Reasoning depth: low/medium/high | `medium` |
33
+ | `--temperature` | Sampling temperature | `1.0` |
34
+ | `--top-p` | Top-p sampling | `1.0` |
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
  ## 💡 What You Get
42
 
 
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)
117
+
118
  ## 📚 Resources
119
 
120
  - [Model: openai/gpt-oss-20b](https://huggingface.co/openai/gpt-oss-20b)
gpt_oss_minimal.py CHANGED
@@ -60,15 +60,37 @@ def main():
60
  parser.add_argument("--model-id", default="openai/gpt-oss-20b", help="Model to use")
61
  parser.add_argument("--max-samples", type=int, help="Limit number of samples")
62
  parser.add_argument(
63
- "--max-new-tokens", type=int, default=1024, help="Max tokens to generate"
64
  )
65
  parser.add_argument(
66
- "--reasoning-effort",
67
- choices=["low", "medium", "high"],
68
  default="medium",
69
- help="Reasoning effort level (default: medium)"
 
 
 
 
 
 
 
 
 
 
 
 
70
  )
71
  args = parser.parse_args()
 
 
 
 
 
 
 
 
 
 
72
 
73
  # Check GPU availability
74
  if not torch.cuda.is_available():
@@ -136,20 +158,21 @@ def main():
136
 
137
  # Apply chat template with reasoning_effort parameter
138
  inputs = tokenizer.apply_chat_template(
139
- messages,
140
- add_generation_prompt=True,
141
- return_tensors="pt",
142
  return_dict=True,
143
- reasoning_effort=args.reasoning_effort # "low", "medium", or "high"
144
  ).to(model.device)
145
 
146
- # Generate
147
  with torch.no_grad():
148
  generated = model.generate(
149
  **inputs,
150
  max_new_tokens=args.max_new_tokens,
151
  do_sample=True,
152
- temperature=0.7,
 
153
  )
154
 
155
  # Decode only the generated part (excluding input)
 
60
  parser.add_argument("--model-id", default="openai/gpt-oss-20b", help="Model to use")
61
  parser.add_argument("--max-samples", type=int, help="Limit number of samples")
62
  parser.add_argument(
63
+ "--max-new-tokens", type=int, help="Max tokens to generate (auto-scales with reasoning effort if not set)"
64
  )
65
  parser.add_argument(
66
+ "--reasoning-effort",
67
+ choices=["low", "medium", "high"],
68
  default="medium",
69
+ help="Reasoning effort level (default: medium)",
70
+ )
71
+ parser.add_argument(
72
+ "--temperature",
73
+ type=float,
74
+ default=1.0,
75
+ help="Sampling temperature (default: 1.0)",
76
+ )
77
+ parser.add_argument(
78
+ "--top-p",
79
+ type=float,
80
+ default=1.0,
81
+ help="Top-p sampling (default: 1.0)",
82
  )
83
  args = parser.parse_args()
84
+
85
+ # Auto-scale max_new_tokens based on reasoning effort if not explicitly set
86
+ if args.max_new_tokens is None:
87
+ if args.reasoning_effort == "high":
88
+ args.max_new_tokens = 2048 # More tokens for detailed reasoning
89
+ elif args.reasoning_effort == "medium":
90
+ args.max_new_tokens = 1024
91
+ else: # low
92
+ args.max_new_tokens = 512
93
+ print(f"Auto-set max_new_tokens={args.max_new_tokens} based on reasoning_effort={args.reasoning_effort}")
94
 
95
  # Check GPU availability
96
  if not torch.cuda.is_available():
 
158
 
159
  # Apply chat template with reasoning_effort parameter
160
  inputs = tokenizer.apply_chat_template(
161
+ messages,
162
+ add_generation_prompt=True,
163
+ return_tensors="pt",
164
  return_dict=True,
165
+ reasoning_effort=args.reasoning_effort, # "low", "medium", or "high"
166
  ).to(model.device)
167
 
168
+ # Generate with user-specified or default parameters
169
  with torch.no_grad():
170
  generated = model.generate(
171
  **inputs,
172
  max_new_tokens=args.max_new_tokens,
173
  do_sample=True,
174
+ temperature=args.temperature,
175
+ top_p=args.top_p,
176
  )
177
 
178
  # Decode only the generated part (excluding input)