# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. # SPDX-License-Identifier: Apache-2.0 video_caption: resources_servers: video_caption: entrypoint: app.py num_workers: 1 judge_model_server: {type: responses_api_models, name: video_caption_judge} # The judge is called via /v1/chat/completions with a JSON-schema response_format per request. judge_chat_completion_create_params: messages: [] max_tokens: 20480 reasoning_effort: low # tokenizer.json of the policy, used only to count caption tokens for the length reward caption_tokenizer_path: ${oc.env:VIDEO_CAPTION_TOKENIZER_PATH,MISSING} judge_endpoint_max_concurrency: 256 judge_max_retries: 2 judge_attempt_timeout_seconds: 1200 min_caption_tokens: 100 max_caption_tokens: 4000 # Visual score combined with the length bonus: recall (default), precision, or f1. # Precision/F1 requires verifier_metadata.reference_caption in each row. visual_reward_type: recall # Opt in from a training recipe. A clear degeneration verdict zeros the total reward. penalize_degeneration: false domain: other verified: false description: Visual event recall, precision, or F1 with length and optional degeneration reward value: Improve coverage and accuracy of visible events in video captions # Video caption reward with a locally served judge (one `local_vllm_model` instance on the Gym's GPU # nodes). Point VIDEO_CAPTION_JUDGE_MODEL at a Hugging Face id or a local checkpoint directory of a # strong instruction/reasoning LLM; the reference runs used a 27B Qwen3-family reasoning model with # thinking enabled at low reasoning effort. See video_caption_api_judge.yaml for an API-served judge. video_caption_judge: responses_api_models: local_vllm_model: entrypoint: app.py num_workers: 1 model: ${oc.env:VIDEO_CAPTION_JUDGE_MODEL,MISSING} return_token_id_information: false uses_reasoning_parser: true # For example, if your model is downloaded at ~/.cache/huggingface/hub/models--Org--Model, hf_home is `~/.cache/huggingface` hf_home: null chat_template_kwargs: {enable_thinking: true, reasoning_effort: low} vllm_serve_env_vars: VLLM_RAY_DP_PACK_STRATEGY: strict OMP_NUM_THREADS: "8" vllm_serve_kwargs: tensor_parallel_size: 1 pipeline_parallel_size: 1 data_parallel_size: 8 api_server_count: 1 dtype: bfloat16 max_model_len: 40960 max_num_seqs: 32 gpu_memory_utilization: 0.92 reasoning_parser: qwen3 video_caption_simple_agent: responses_api_agents: simple_agent: entrypoint: app.py max_steps: 1 resources_server: {type: resources_servers, name: video_caption} model_server: {type: responses_api_models, name: policy_model} datasets: - name: example type: example jsonl_fpath: resources_servers/video_caption/data/example.jsonl num_repeats: 1 video_caption_environment_server: environment_servers: single_agent_turn_legacy: entrypoint: app.py agent_server: type: responses_api_agents name: video_caption_simple_agent resources_server: type: resources_servers name: video_caption resources_tool_transports: - direct_http