Multi-Domain

conf/base.yaml

@@ -141,4 +141,3 @@ wandb:
   wandb_dir: null
   # Comma-separated list of keywords to tag the run.
   tags: []
-

conf/coding.yaml

@@ -0,0 +1,47 @@
+defaults:
+  - base
+  - _self_
+
+actor:
+  rollout_policy: pipelinerl.domains.coding.generate_coding_rollout
+  # Empty system prompt - the problems already include instructions like:
+  # "Read the inputs from stdin...Enclose your code within ```python delimiters."
+  # Adding extra instructions can conflict with the model's native prompt format.
+  system_prompt: ""
+  task_template: |-
+    {task}
+
+    Write your solution in Python. Be concise. Output only the code in a single ```python``` block.
+  task_prompt: ""
+  ensure_boxed_answers: false
+
+  # SandboxFusion verification endpoint (override via SANDBOX_ENDPOINT env var)
+  sandbox_endpoint: ${oc.env:SANDBOX_ENDPOINT,http://127.0.0.1:8080}
+  sandbox_timeout: 10.0
+  max_tests_per_problem: 5
+
+# Use combined TACO + APPS dataset
+dataset_loader: pipelinerl.domains.coding.dataset.load_problems
+dataset_loader_params:
+  taco_split: train
+  apps_split: train
+  subset: train  # "train", "test", or "all"
+  train_ratio: 0.9
+  # Filter out HARD and VERY_HARD from TACO (APPS kept as-is)
+  taco_excluded_difficulties:
+    - HARD
+    - VERY_HARD
+  # Limit test cases per problem (aligned with median test counts)
+  max_tests_per_problem: 5
+  max_examples: null
+  seed: 42
+  huggingface_token: ${oc.env:HF_TOKEN, null}
+
+train_dataset_names:
+  - taco
+  - apps
+
+test_dataset_names:
+  - livecodebench_v5
+
+environment_key: coding

conf/debug/multi_domain.yaml

@@ -0,0 +1,33 @@
+# @package _global_
+defaults:
+  - /base
+  - /domain_rollouts@domain_rollouts: base
+  - override /rewards: success_and_format
+  - _self_
+
+actor:
+  rollout_policy: pipelinerl.domains.dispatcher.generate_multidomain_rollout
+  llm_max_rollouts: 2
+  rollout_workers: 1
+  domain_rollouts:
+    math: ${domain_rollouts.math}
+    guessing: ${domain_rollouts.guessing}
+    coding: ${domain_rollouts.coding}
+
+dataset_loader: pipelinerl.domains.multidomain.load_problems
+train_dataset_names:
+  - math_debug
+  - guessing_debug
+  - coding_debug
+test_dataset_names:
+  - math_debug
+  - coding_debug
+
+environments:
+  - key: math
+    mode: remote
+    replicas_per_actor: ${world.env_replicas_per_actor}
+    _target_: pipelinerl.domains.math.MathEnvironment
+
+world:
+  env_replicas_per_actor: 1

conf/domain_mix/base.yaml

@@ -0,0 +1,8 @@
+# @package actor.domain_mix
+# Default domain mix weights for multi-domain training
+
+math: 0.35
+coding: 0.25
+logic: 0.15
+ifeval: 0.15
+fn_calling: 0.10

conf/domain_rollouts/base.yaml

@@ -0,0 +1,11 @@
+# Mapping between domain identifiers and rollout callables.
+math: pipelinerl.domains.math.generate_math_rollout
+guessing: pipelinerl.domains.guessing.generate_guessing_rollout
+counting: pipelinerl.domains.counting.generate_counting_rollout
+miniwob: pipelinerl.domains.miniwob.rollouts.generate_miniwob_rollout
+chartqa: pipelinerl.domains.chartqa.generate_chartqa_rollout
+coding: pipelinerl.domains.coding.generate_coding_rollout
+livecodebench: pipelinerl.domains.coding.generate_coding_rollout
+fn_calling: pipelinerl.domains.fn_calling.generate_fn_calling_rollout
+logic: pipelinerl.domains.logic.generate_logic_rollout
+ifeval: pipelinerl.domains.ifeval.generate_ifeval_rollout

conf/fn_calling.yaml

@@ -0,0 +1,53 @@
+defaults:
+  - base
+  - _self_
+
+actor:
+  rollout_policy: pipelinerl.domains.fn_calling.generate_fn_calling_rollout
+  # System prompt is intentionally empty - the chat template will inject
+  # tool instructions when tools are present in the request
+  system_prompt: ""
+  task_template: |-
+    {task}
+  task_prompt: ""
+  ensure_boxed_answers: false
+
+dataset_loader: pipelinerl.domains.fn_calling.dataset.load_problems
+dataset_loader_params:
+  # BFCL v4 test categories (single-turn) - ~2500 samples total
+  categories:
+    # Simple (single function call)
+    - simple_python      # 400
+    - simple_java        # 100
+    - simple_javascript  # 50
+    # Complex (multiple calls)
+    - multiple           # 200
+    - parallel           # 200
+    - parallel_multiple  # 200
+    # Live (real-world APIs)
+    - live_simple        # 258
+    - live_multiple      # 1053
+    - live_parallel      # 16
+    - live_parallel_multiple  # 24
+  max_examples_per_category: null
+  max_examples: null
+  subset: train  # "train" or "test" (90/10 split by default)
+  train_ratio: 0.9
+  seed: 42
+
+train_dataset_names:
+  - fn_calling
+
+test_dataset_names:
+  - fn_calling_test
+
+environments:
+  - key: fn_calling
+    mode: remote
+    replicas_per_actor: ${world.env_replicas_per_actor}
+    _target_: pipelinerl.domains.fn_calling.BFCLEnvironment
+
+environment_key: fn_calling
+
+world:
+  env_replicas_per_actor: 1

conf/ifeval.yaml

@@ -0,0 +1,47 @@
+defaults:
+  - base
+  - _self_
+
+actor:
+  rollout_policy: pipelinerl.domains.ifeval.generate_ifeval_rollout
+  system_prompt: ""
+  task_template: |-
+    {task}
+
+    Be concise and direct in your response. Do not add unnecessary elaboration.
+  task_prompt: ""
+  ensure_boxed_answers: false
+
+  # IFEval verification options
+  ifeval_partial_credit: true  # If true, give partial reward for some constraints met
+
+dataset_loader: pipelinerl.domains.ifeval.dataset.load_problems
+dataset_loader_params:
+  # AllenAI IF_multi_constraints_upto5 (95k training samples with up to 5 constraints)
+  max_examples: null
+  max_constraints: null  # Optionally limit max constraints per sample
+  seed: 42
+  huggingface_token: ${oc.env:HF_TOKEN, null}
+  # Train/test split: ~550 held-out for testing (similar to Google IFEval size)
+  test_size: 550
+  # Filter out prompts exceeding this token count (max_model_len - max_tokens - buffer)
+  # With max_model_len=48000 and max_tokens=4096: 48000 - 4096 - 1000 = 42904
+  max_prompt_tokens: 42000
+
+train_dataset_names:
+  - ifeval
+
+test_dataset_names:
+  - ifeval_test
+
+environments:
+  - key: ifeval
+    mode: remote
+    replicas_per_actor: ${world.env_replicas_per_actor}
+    _target_: pipelinerl.domains.ifeval.IFEvalEnvironment
+
+world:
+  env_replicas_per_actor: 1
+
+# IFEval verification via RPC to IFEvalEnvironment server
+environment_key: ifeval

conf/logic.yaml

@@ -0,0 +1,49 @@
+defaults:
+  - base
+  - _self_
+
+actor:
+  rollout_policy: pipelinerl.domains.logic.generate_logic_rollout
+  system_prompt: ""
+  task_template: |-
+    {task}
+
+  task_prompt: ""
+  ensure_boxed_answers: false
+
+dataset_loader: pipelinerl.domains.logic.dataset.load_problems
+dataset_loader_params:
+  # PrimeIntellect INTELLECT-3-RL logic domain (11.6k samples, 87 task types)
+  dataset_id: PrimeIntellect/INTELLECT-3-RL
+  dataset_config: logic
+  split: train
+  subset: train  # "train" or "test"
+  # Fixed test size similar to IFEval (~550 held-out samples)
+  test_size: 550
+  # Difficulty filtering using model solve rates
+  min_difficulty: 0.0
+  max_difficulty: 1.0
+  difficulty_column: avg@16_qwen3_4b_instruct_2507
+  # Skip certain task types (as recommended by i3-logic)
+  tasks_to_skip: [arc_agi, arc_agi_2, buggy_tables, dyck_language, dyck_language_errors, dyck_language_reasoning_errors]
+  max_examples: null
+  seed: 42
+  huggingface_token: ${oc.env:HF_TOKEN, null}
+
+train_dataset_names:
+  - logic
+
+test_dataset_names:
+  - logic_test
+
+environments:
+  - key: logic
+    mode: remote
+    replicas_per_actor: ${world.env_replicas_per_actor}
+    _target_: pipelinerl.domains.logic.LogicEnvironment
+
+world:
+  env_replicas_per_actor: 1
+
+# Logic verification via RPC to LogicEnvironment server
+environment_key: logic

conf/math.yaml

@@ -5,10 +5,13 @@ defaults:
 actor:
   rollout_policy: pipelinerl.domains.math.generate_math_rollout
   system_prompt: Please reason step by step, and put your final answer within \boxed{}.
-  task_template: |-
-    {task}
-environment:
-  _target_: pipelinerl.domains.math.MathEnvironment
+  task_template: "{task}"
+  task_prompt: ""
+environments:
+  - key: math
+    mode: remote
+    _target_: pipelinerl.domains.math.MathEnvironment
+environment_key: math
 dataset_loader: pipelinerl.domains.math.load_datasets
 
 finetune:
@@ -32,4 +35,4 @@ train_dataset_names:
 test_dataset_names:
   - aime_2024
   - amc_2023
-  - math_500
+  - math_500