feat: prime tau2 synthetic training environment

environments/prime_tau2_synth/README.md

@@ -0,0 +1,47 @@
+# Description
+
+This environment is for training on τ²-bench style tasks with custom synthetic data via [tau2-synth](https://app.primeintellect.ai/dashboard/environments/prime/tau2-synth) on Prime Intellect Environments hub through verifiers integration. See tau2-synth for the full description and configuration details.
+
+Domains: `library`, `fitness_gym`, `tech_support`, `telecom`, `cloud_incident_response`, `daily_planner`, `ev_charging_support`.
+
+For additional details on the prime verifiers integration in NeMo Gym, see [responses_api_agents/verifiers_agent/README.md](../../responses_api_agents/verifiers_agent/README.md).
+
+## Install Gym
+
+```
+git clone https://github.com/NVIDIA-NeMo/Gym
+cd Gym
+uv venv; source .venv/bin/activate; uv sync
+```
+
+## Test tau2-synth example
+
+First set `env.yaml`, for example for a vLLM served model:
+```
+policy_base_url: "http://localhost:8000/v1"
+policy_api_key: EMPTY
+policy_model_name: "Qwen/Qwen3-4B-Instruct-2507"
+```
+
+```
+# start nemo gym servers
+ng_run "+config_paths=[environments/prime_tau2_synth/config.yaml,responses_api_models/vllm_model/configs/vllm_model.yaml]"
+
+# generate a rollout
+ng_collect_rollouts \
+    +agent_name=prime_tau2_synth_agent \
+    +input_jsonl_fpath=environments/prime_tau2_synth/data/example.jsonl \
+    +output_jsonl_fpath=environments/prime_tau2_synth/data/example-rollouts.jsonl \
+    +limit=1
+
+# view the rollout
+tail -n 1 environments/prime_tau2_synth/data/example-rollouts.jsonl | jq | less
+```
+
+# Licensing information
+Code: Apache 2.0
+Data: N/A
+
+Dependencies
+- nemo_gym: Apache 2.0
+- verifiers: Apache 2.0

environments/prime_tau2_synth/config.yaml

@@ -0,0 +1,27 @@
+prime_tau2_synth_agent:
+  responses_api_agents:
+    prime_tau2_synth_agent:
+      entrypoint: app.py
+      domain: agent
+      description: tau2-bench style training with custom synthetic domains via Prime Intellect verifiers.
+      value: Improve multi-domain agentic customer service and tool use capabilities.
+      model_server:
+        type: responses_api_models
+        name: policy_model
+      model_name: ""
+      vf_env_id: tau2-synth
+      vf_env_args:
+        domain: library
+        user_model: gpt-4.1
+        max_steps: 200
+        max_errors: 10
+      group_size: 1
+      max_concurrent_generation: -1
+      max_concurrent_scoring: -1
+      max_tokens: 16000
+      temperature: 1.0
+      top_p: 1.0
+      datasets:
+        - name: example
+          type: example
+          jsonl_fpath: environments/prime_tau2_synth/data/example.jsonl

environments/prime_tau2_synth/data/.gitignore

@@ -0,0 +1,5 @@
+*train.jsonl
+*validation.jsonl
+*train_prepare.jsonl
+*validation_prepare.jsonl
+*example_prepare.jsonl

environments/prime_tau2_synth/data/example.jsonl

@@ -0,0 +1 @@
+{"task_idx": 0, "vf_env_id": "tau2-synth", "responses_create_params": {"input": []}}

environments/prime_tau2_synth/prepare.py

@@ -0,0 +1,82 @@
+# Copyright (c) 2026, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Generate a Gym JSONL for the tau2-synth verifiers environment.
+
+Loads the tau2-synth env for a given domain and writes one row per task.
+
+Requires the agent venv (or a venv with verifiers + tau2-synth installed):
+    pip install 'verifiers @ git+https://github.com/PrimeIntellect-ai/verifiers.git@v0.1.14' \\
+        --extra-index-url https://hub.primeintellect.ai/prime/simple/
+    pip install 'tau2 @ git+https://github.com/eligotts/tau2-bench.git@4839dd6' tau2-synth==0.2.0
+
+Usage:
+    python environments/prime_tau2_synth/prepare.py --domain library --output environments/prime_tau2_synth/data/library_train.jsonl
+"""
+
+import argparse
+import json
+from pathlib import Path
+
+
+VF_ENV_ID = "tau2-synth"
+DOMAINS = [
+    "library",
+    "fitness_gym",
+    "tech_support",
+    "telecom",
+    "cloud_incident_response",
+    "daily_planner",
+    "ev_charging_support",
+]
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--domain", default="library", choices=DOMAINS)
+    parser.add_argument("--split", default="train", help="verifiers env split (typically 'train' or 'test')")
+    parser.add_argument("--limit", type=int, default=None)
+    parser.add_argument("--output", required=True)
+    args = parser.parse_args()
+
+    try:
+        import verifiers as vf
+    except ImportError:
+        raise ImportError("Install verifiers + tau2-synth first; see module docstring.")
+
+    env = vf.load_environment(VF_ENV_ID, domain=args.domain)
+    dataset = env.get_dataset(args.split)
+    rows = list(dataset)
+    if args.limit is not None:
+        rows = rows[: args.limit]
+
+    output_path = Path(args.output)
+    output_path.parent.mkdir(parents=True, exist_ok=True)
+    with output_path.open("w") as f:
+        for idx, row in enumerate(rows):
+            out = {
+                "task_idx": idx,
+                "vf_env_id": VF_ENV_ID,
+                "domain": args.domain,
+                "responses_create_params": {
+                    "input": [{"role": "user", "content": row.get("question", row.get("prompt", ""))}]
+                },
+                **{k: v for k, v in row.items() if k not in {"responses_create_params"}},
+            }
+            f.write(json.dumps(out) + "\n")
+
+    print(f"Wrote {len(rows)} rows to {output_path}")
+
+
+if __name__ == "__main__":
+    main()

responses_api_agents/prime_tau2_synth_agent/app.py

@@ -0,0 +1,18 @@
+# Copyright (c) 2026, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from responses_api_agents.verifiers_agent.app import VerifiersAgent
+
+
+if __name__ == "__main__":
+    VerifiersAgent.run_webserver()

responses_api_agents/prime_tau2_synth_agent/requirements.txt

@@ -0,0 +1,5 @@
+-e nemo-gym[dev] @ ../../
+verifiers @ git+https://github.com/PrimeIntellect-ai/verifiers.git@v0.1.14
+--extra-index-url https://hub.primeintellect.ai/prime/simple/
+tau2 @ git+https://github.com/eligotts/tau2-bench.git@4839dd6
+tau2-synth==0.2.0