Skip to content

Providers and models

Run AdaMAST generation and judging on the provider you already use. Both workflows share one provider-neutral text interface, so the prompts and output validation stay the same when the provider changes.

Run this first

Commands on this page read the bundled examples. Create them in the directory you are working from with python -m adamast.examples.

馃З Supported providers

Provider flag Install extra Credential environment Model environment
openai Included with pip install adamast OPENAI_API_KEY OPENAI_MODEL
anthropic [anthropic] ANTHROPIC_API_KEY ANTHROPIC_MODEL
google [google] GEMINI_API_KEY or GOOGLE_API_KEY GEMINI_MODEL or GOOGLE_MODEL
bedrock [bedrock] AWS bearer token or normal AWS credential chain BEDROCK_MODEL_ID

Select a provider explicitly with --provider or ADAMAST_PROVIDER. Also pass --model or set the provider's model environment variable; only OpenAI ships a package default model.

馃煝 OpenAI

pip install adamast
export OPENAI_API_KEY="..."

adamast generate \
  --provider openai \
  --model gpt-5-nano \
  --traces adamast-examples/traces.jsonl \
  --output ./taxonomy-run

Note

If neither --model nor OPENAI_MODEL is set, the current package defaults to gpt-5-nano.

馃煟 Anthropic

pip install "adamast[anthropic]"
export ANTHROPIC_API_KEY="..."
export ANTHROPIC_MODEL="YOUR_MODEL_ID"

adamast generate \
  --provider anthropic \
  --traces adamast-examples/traces.jsonl \
  --output ./taxonomy-run

馃數 Google

pip install "adamast[google]"
export GEMINI_API_KEY="..."
export GEMINI_MODEL="YOUR_MODEL_ID"

adamast generate \
  --provider google \
  --traces adamast-examples/traces.jsonl \
  --output ./taxonomy-run

GOOGLE_API_KEY and GOOGLE_MODEL are accepted aliases.

馃煚 AWS Bedrock

pip install "adamast[bedrock]"
export AWS_REGION="us-east-1"
export BEDROCK_MODEL_ID="YOUR_BEDROCK_MODEL_ID"

adamast generate \
  --provider bedrock \
  --traces adamast-examples/traces.jsonl \
  --output ./taxonomy-run

AdaMAST uses the Bedrock Runtime Converse API. Authentication can come from AWS_BEARER_TOKEN_BEDROCK or the normal boto3 chain: environment credentials, shared configuration, an AWS profile, container credentials, or an instance role.

Choose a profile and region explicitly when needed:

adamast generate \
  --provider bedrock \
  --model YOUR_BEDROCK_MODEL_ID \
  --aws-profile research \
  --aws-region us-west-2 \
  --traces adamast-examples/traces.jsonl \
  --output ./taxonomy-run

馃尡 Environment-only configuration

export ADAMAST_PROVIDER="anthropic"
export ANTHROPIC_API_KEY="..."
export ANTHROPIC_MODEL="YOUR_MODEL_ID"

adamast generate --traces adamast-examples/traces.jsonl --output ./taxonomy-run

Explicit CLI values take precedence over model environment variables.

鈴憋笍 Output and timeout controls

--max-output-tokens caps the output of each model call. The default is 8192 for both generation and judging; pass the flag only when you want a different cap; here, lowering it to 4096:

adamast judge \
  --provider google \
  --model YOUR_MODEL_ID \
  --max-output-tokens 4096 \
  --taxonomy adamast-examples/taxonomy.sample.json \
  --traces adamast-examples/traces.jsonl

No silent fallback

Provider request errors stop the workflow. AdaMAST does not silently switch to another provider or model.

馃攼 Credential safety

  • Put credentials in the provider's environment or standard credential store, never in trace files or AdaMAST JSON artifacts.
  • Redact secrets from trajectories before generation or judging.
  • Use least-privilege AWS credentials that allow only the required Bedrock model actions.
  • Treat model IDs as experiment inputs and record them in reproducible runs.

鉃★笍 Continue with