rnow 0.2.4__py3-none-any.whl

rnow/templates/rl-single/config.yml

@@ -0,0 +1,27 @@
+project_id: ""
+project_name: "OpenMathReasoning"
+dataset_id: ""
+dataset_name: "math-problems"
+dataset_type: rl
+organization_id: ""
+data:
+  train_file: train.jsonl
+  batch_size: 32
+  group_size: 16
+model:
+  path: Qwen/Qwen3-8B
+  qlora_rank: 32
+  name: "Math Reasoning Model"
+  description: "RL model for mathematical reasoning"
+algorithm:
+  loss_fn: ppo
+  adv_estimator: grpo
+  kl_penalty_coef: 0.01
+rollout:
+  max_turns: 1
+  max_tokens: 16384
+  termination_policy: last_tool
+trainer:
+  num_epochs: 4
+  learning_rate: 0.0001
+  save_step: 333

rnow/templates/rl-single/requirements.txt

@@ -0,0 +1 @@
+math-verify==0.5.0

rnow/templates/rl-single/rewards.py

@@ -0,0 +1,14 @@
+from math_verify import LatexExtractionConfig, parse, verify
+
+from rnow.core import RewardArgs, reward
+
+
+@reward
+def accuracy(args: RewardArgs, messages: list) -> float:
+    gold = parse(args.metadata["expected_answer"])
+    pred = parse(
+        messages[-1]["content"], extraction_config=[LatexExtractionConfig(boxed_match_priority=0)]
+    )
+    if not pred:
+        return 0.0
+    return 1.0 if verify(gold, pred) else 0.0