Cloud Parameter Equivalence¶

This page maps parameters across all supported cloud platforms so you can use the same tuning strategy everywhere.

Core Parameters¶

Temperature¶

Controls randomness/creativity of responses.

Platform	Parameter	Range	Default	Notes
Azure OpenAI	`temperature`	0-2	1.0	Can exceed 1.0
OpenAI	`temperature`	0-2	1.0	Can exceed 1.0
Vertex AI	`temperature`	0-2	1.0	Can exceed 1.0
Bedrock (Claude)	`temperature`	0-1	1.0	Capped at 1.0
Bedrock (Llama)	`temperature`	0-1	0.5	Capped at 1.0

Migration tip: If tuning temperature for Azure/OpenAI/Vertex and want to use Bedrock, divide by 2 if > 1.0.

Top-p (Nucleus Sampling)¶

Filters candidate tokens via cumulative probability.

Platform	Parameter	Range	Default	Notes
Azure OpenAI	`top_p`	0-1	1.0	Pure nucleus
OpenAI	`top_p`	0-1	1.0	Pure nucleus
Vertex AI	`top_p`	0-1	1.0	Pure nucleus
Bedrock (Anthropic)	`top_p`	0-1	0.999	Very close to 1 by default
Bedrock (Meta/Mistral)	`top_p`	0-1	0.9	Slightly tighter

Migration tip: All platforms support 0-1 range identically.

Top-k¶

Keeps top-k most likely tokens.

Platform	Parameter	Range	Default	Notes
Azure OpenAI	N/A	—	—	Use top_p instead
OpenAI	N/A	—	—	Use top_p instead
Vertex AI	`top_k`	1-40	N/A	Optional, works with top_p
Bedrock (Anthropic)	`top_k`	0-500	N/A	Optional, not commonly used
Bedrock (Meta)	`top_k`	0-500	N/A	Optional

Migration tip: Azure/OpenAI don't support top_k; use top_p for similar effect.

Max Tokens / Max Output¶

Limits response length.

Platform	Parameter	Notes
Azure OpenAI	`max_completion_tokens`	Renamed from `max_tokens` in v2024-10-21+
OpenAI	`max_tokens`	Standard parameter
Vertex AI	`max_output_tokens`	Vertex-specific naming
Bedrock	`max_tokens`	Standard parameter

Migration tip: Always cap max tokens appropriately for your use case.

Stop Sequences¶

Tells model when to stop generating.

Platform	Parameter	Format	Notes
Azure OpenAI	`stop`	List of strings	Up to 4 stop sequences
OpenAI	`stop`	List of strings	Up to 4 stop sequences
Vertex AI	`stop_sequences`	List of strings	Variable limit
Bedrock	`stop_sequences`	List of strings	Variable limit

Migration tip: Same concept, slightly different parameter names.

Advanced Parameters¶

Frequency Penalty¶

Reduces likelihood of repeated tokens.

Platform	Supported	Range	Notes
Azure OpenAI	✅ Yes	-2 to 2	Called `frequency_penalty`
OpenAI	✅ Yes	-2 to 2	Called `frequency_penalty`
Vertex AI	❌ No	—	Use top_p instead
Bedrock	❌ No	—	Not directly supported

Presence Penalty¶

Penalizes new tokens that haven't appeared yet.

Platform	Supported	Range	Notes
Azure OpenAI	✅ Yes	-2 to 2	Called `presence_penalty`
OpenAI	✅ Yes	-2 to 2	Called `presence_penalty`
Vertex AI	❌ No	—	Not available
Bedrock	❌ No	—	Not available

Seed / Determinism¶

Makes responses reproducible with same seed.

Platform	Parameter	Reproducible	Notes
Azure OpenAI	`seed`	✅ Yes	Guarantees determinism (with caveats)
OpenAI	`seed`	✅ Yes	Best-effort determinism
Vertex AI	`seed`	⚠️ Partial	Reduces variability but not guaranteed
Bedrock	N/A	❌ No	No native seed parameter

Quick Migration Guide¶

From Azure OpenAI → Bedrock (Claude)¶

# Azure
{
    "temperature": 0.8,
    "top_p": 0.95,
    "presence_penalty": 0.2,
    "frequency_penalty": 0.0,
    "max_completion_tokens": 256
}

# Bedrock equivalent (what to keep/drop)
{
    "temperature": 0.8,      # Keep as-is
    "top_p": 0.95,            # Keep as-is
    # presence_penalty: drop (not supported)
    # frequency_penalty: drop (not supported)
    "max_tokens": 256         # Rename from max_completion_tokens
}

From OpenAI → Vertex AI¶

# OpenAI
{
    "temperature": 1.0,
    "top_p": 0.9,
    "frequency_penalty": 0.1,
    "max_tokens": 500,
    "stop": ["\n\nUser:"]
}

# Vertex equivalent
{
    "temperature": 1.0,         # Keep as-is
    "top_p": 0.9,                # Keep as-is
    # frequency_penalty: drop (not supported)
    "max_output_tokens": 500,   # Rename
    "stop_sequences": ["\n\nUser:"]  # Rename
}

From Bedrock (Llama) → Azure OpenAI¶

# Bedrock
{
    "temperature": 0.5,
    "top_p": 0.9,
    "top_k": 40,
    "max_tokens": 256
}

# Azure equivalent
{
    "temperature": 0.5 * 2,      # Llama temp is ~half of Azure
    "top_p": 0.9,                 # Keep as-is
    # top_k: drop (not supported, use top_p)
    "max_completion_tokens": 256  # Rename
}

Testing Across Platforms¶

When moving parameters from one platform to another:

Keep same temperature/top_p values if ranges overlap
Drop unsupported parameters (e.g., frequency_penalty for Bedrock)
Test on new platform with 2-3 small samples first
Compare outputs for consistency
Adjust incrementally if behavior differs

Example:

# Test same prompt on 3 platforms
python3 scripts/cloud/azure_sweep.py --preset summarization --trials 1
python3 scripts/cloud/openai_sweep.py --preset summarization --trials 1
python3 scripts/cloud/bedrock_sweep.py --preset summarization --trials 1

# Compare results
python3 scripts/analyze_results.py --compare-platforms

See also: Platform Setup Guides