Add RL training configs, scripts, and environment updates

.gitignore

@@ -1,5 +1,17 @@
 outputs/
 
+# Prime-RL (separate repo with trained models)
+prime-rl/
+
+# CSV data exports
+*.csv
+
+# Generated plots
+plots/
+
+# Eval results
+eval_results/
+
 # Byte-compiled / optimized / DLL files
 __pycache__/
 *.py[codz]

configs/medcalc_bench_8b_nothink_rl.toml

@@ -0,0 +1,72 @@
+inference_gpu_ids = [0, 1, 2, 3, 4, 5]
+trainer_gpu_ids = [6, 7]
+
+max_steps = 300
+seq_len = 4096 
+
+[wandb]
+project = "medcalc-bench-verified-8b"
+name = "medcalc-bench-verified-8b-nothink"
+
+[ckpt]
+# Checkpoint at the end of training
+
+[model]
+name = "Qwen/Qwen3-8B-Instruct-2507"
+
+[orchestrator]
+batch_size = 128  # Reduced for 8B model
+rollouts_per_example = 16
+
+[[orchestrator.env]]
+id = "medcalc_bench"
+
+# Environment arguments for MedCalc-Bench
+[orchestrator.env.args]
+one_shot = false
+add_python_tool = false
+use_think = false  # No thinking tags
+answer_format = "xml"
+use_verified_dataset = true  # Use MedCalc-Bench-Verified
+
+[orchestrator.sampling]
+max_tokens = 1024
+temperature = 1.0
+
+# Evaluation configuration
+[orchestrator.eval]
+interval = 10
+eval_base_model = true
+
+[[orchestrator.eval.env]]
+id = "medcalc_bench"
+num_examples = 300
+rollouts_per_example = 1
+
+[orchestrator.eval.env.args]
+one_shot = false
+add_python_tool = false
+use_think = false  # No thinking tags
+answer_format = "xml"
+use_verified_dataset = true  # Use MedCalc-Bench-Verified
+
+[orchestrator.eval.sampling]
+temperature = 0.0
+max_tokens = 1024
+
+[trainer]
+# Default trainer config (GRPO)
+
+[trainer.model.lora]
+rank = 32
+alpha = 64.0
+
+[trainer.optim]
+lr = 1e-5
+
+[inference]
+# Inference server config
+
+[inference.parallel]
+tp = 2
+dp = 3

configs/medcalc_bench_rl.toml

@@ -0,0 +1,73 @@
+# MedCalc-Bench RL Training Configuration
+# Clinical calculator reasoning with numeric/date outputs
+
+inference_gpu_ids = [0, 1, 2, 3, 4, 5]
+trainer_gpu_ids = [6, 7]
+
+max_steps = 300
+seq_len = 4096
+
+[wandb]
+project = "medcalc-bench-rl-1"
+name = "medcalc-bench-rl-1"
+
+[ckpt]
+# Checkpoint at the end of training
+
+[model]
+name = "Qwen/Qwen3-4B-Instruct-2507"
+
+[orchestrator]
+batch_size = 512
+rollouts_per_example = 16
+
+[[orchestrator.env]]
+id = "medcalc_bench"
+
+# Environment arguments for MedCalc-Bench
+[orchestrator.env.args]
+one_shot = false          # Zero-shot prompting
+add_python_tool = false   # No code execution
+use_think = true          # Enable <think> reasoning tags
+answer_format = "xml"     # Use XML format for answers
+
+[orchestrator.sampling]
+max_tokens = 2048
+temperature = 1.0
+
+# Evaluation configuration
+[orchestrator.eval]
+interval = 10  # Evaluate every 10 steps
+eval_base_model = true  # Also evaluate the base model at step 0
+
+[[orchestrator.eval.env]]
+id = "medcalc_bench"
+num_examples = 300 
+rollouts_per_example = 1
+
+[orchestrator.eval.env.args]
+one_shot = false
+add_python_tool = false
+use_think = true
+answer_format = "xml"
+
+[orchestrator.eval.sampling]
+temperature = 0.0  # Greedy sampling for evaluation
+max_tokens = 2048
+
+[trainer]
+# Default trainer config (GRPO)
+
+[trainer.model.lora]
+rank = 32
+alpha = 64.0
+
+[trainer.optim]
+lr = 1e-5  # LoRA typically benefits from higher LR
+
+[inference]
+# Inference server config
+
+[inference.parallel]
+tp = 2  # Split model across 2 GPUs for faster per-query latency
+dp = 3  # 3 parallel instances

configs/medcalc_bench_short_rl.toml

@@ -0,0 +1,72 @@
+inference_gpu_ids = [0, 1, 2, 3, 4, 5]
+trainer_gpu_ids = [6, 7]
+
+max_steps = 300
+seq_len = 4096 
+
+[wandb]
+project = "medcalc-bench-verified"
+name = "medcalc-bench-verified-2048tok"
+
+[ckpt]
+# Checkpoint at the end of training
+
+[model]
+name = "Qwen/Qwen3-4B-Instruct-2507"
+
+[orchestrator]
+batch_size = 256  
+rollouts_per_example = 16
+
+[[orchestrator.env]]
+id = "medcalc_bench"
+
+# Environment arguments for MedCalc-Bench
+[orchestrator.env.args]
+one_shot = false
+add_python_tool = false
+use_think = true
+answer_format = "xml"
+use_verified_dataset = true  # Use MedCalc-Bench-Verified
+
+[orchestrator.sampling]
+max_tokens = 1024
+temperature = 1.0
+
+# Evaluation configuration
+[orchestrator.eval]
+interval = 10
+eval_base_model = true
+
+[[orchestrator.eval.env]]
+id = "medcalc_bench"
+num_examples = 300
+rollouts_per_example = 1
+
+[orchestrator.eval.env.args]
+one_shot = false
+add_python_tool = false
+use_think = true
+answer_format = "xml"
+use_verified_dataset = true  # Use MedCalc-Bench-Verified
+
+[orchestrator.eval.sampling]
+temperature = 0.0
+max_tokens = 1024
+
+[trainer]
+# Default trainer config (GRPO)
+
+[trainer.model.lora]
+rank = 32
+alpha = 64.0
+
+[trainer.optim]
+lr = 1e-5
+
+[inference]
+# Inference server config
+
+[inference.parallel]
+tp = 2
+dp = 3

configs/medcalc_bench_tools_rl.toml

@@ -0,0 +1,77 @@
+# MedCalc-Bench RL Training Configuration (with Python Tool)
+# Clinical calculator reasoning with code execution
+
+inference_gpu_ids = [0, 1, 2, 3, 4, 5]
+trainer_gpu_ids = [6, 7]
+
+max_steps = 300
+seq_len = 8192  # Longer for multi-turn tool use
+
+[wandb]
+project = "medcalc-bench-tools"
+name = "medcalc-bench-tools-rl"
+
+[ckpt]
+# Checkpoint at the end of training
+
+[model]
+name = "Qwen/Qwen3-4B-Instruct-2507"
+
+[orchestrator]
+batch_size = 256  # Smaller batch for multi-turn (more tokens per example)
+rollouts_per_example = 8
+
+[[orchestrator.env]]
+id = "medcalc_bench"
+
+# Environment arguments for MedCalc-Bench with tools
+[orchestrator.env.args]
+one_shot = false
+add_python_tool = true    # Enable Python tool for calculations
+use_think = true
+answer_format = "xml"
+max_turns = 10            # Allow multiple tool calls
+use_verified_dataset = true  # Use MedCalc-Bench-Verified
+
+[orchestrator.sampling]
+max_tokens = 2048
+temperature = 1.0
+
+# Evaluation configuration
+[orchestrator.eval]
+interval = 10
+eval_base_model = true
+
+[[orchestrator.eval.env]]
+id = "medcalc_bench"
+num_examples = 300
+rollouts_per_example = 1
+
+[orchestrator.eval.env.args]
+one_shot = false
+add_python_tool = true    # Tools enabled for eval too
+use_think = true
+answer_format = "xml"
+max_turns = 10
+use_verified_dataset = true  # Use MedCalc-Bench-Verified
+
+[orchestrator.eval.sampling]
+temperature = 0.0
+max_tokens = 2048
+
+[trainer]
+# Default trainer config (GRPO)
+
+[trainer.model.lora]
+rank = 32
+alpha = 64.0
+
+[trainer.optim]
+lr = 1e-5
+
+[inference]
+# Inference server config
+
+[inference.parallel]
+tp = 2
+dp = 3

configs/medcasereasoning_rl.toml

@@ -0,0 +1,72 @@
+# MedCaseReasoning RL Training Configuration
+# Medical diagnosis from case presentations using LLM-as-a-Judge
+
+inference_gpu_ids = [0, 1, 2, 3, 4, 5]
+trainer_gpu_ids = [6, 7]
+
+max_steps = 300
+seq_len = 2048  # Prompts are shorter (max ~1034 tokens)
+
+[wandb]
+project = "medcasereasoning-rl"
+name = "medcasereasoning-rl"
+
+[ckpt]
+# Checkpoint at the end of training
+
+[model]
+name = "Qwen/Qwen3-4B-Instruct-2507"
+
+[orchestrator]
+batch_size = 512
+rollouts_per_example = 16
+
+[[orchestrator.env]]
+id = "medcasereasoning"
+
+# Environment arguments - uses LLM-as-a-Judge
+[orchestrator.env.args]
+judge_model = "gpt-5-nano"
+reasoning_effort = "low"
+# judge_base_url = "http://localhost:8001/v1"  # Uncomment for local judge
+# judge_api_key = "your-api-key"  # Optional, uses OPENAI_API_KEY by default
+
+[orchestrator.sampling]
+max_tokens = 1024
+temperature = 1.0
+
+# Evaluation configuration
+[orchestrator.eval]
+interval = 10
+eval_base_model = true
+
+[[orchestrator.eval.env]]
+id = "medcasereasoning"
+num_examples = 100  # Smaller eval set (LLM judge is slow/expensive)
+rollouts_per_example = 1
+
+[orchestrator.eval.env.args]
+judge_model = "gpt-5-nano"
+reasoning_effort = "low"
+eval_split = "test"  # Use test set for evaluation
+
+[orchestrator.eval.sampling]
+temperature = 0.0
+max_tokens = 1024
+
+[trainer]
+# Default trainer config (GRPO)
+
+[trainer.model.lora]
+rank = 32
+alpha = 64.0
+
+[trainer.optim]
+lr = 1e-5
+
+[inference]
+# Inference server config
+
+[inference.parallel]
+tp = 2
+dp = 3