OpenRLHF/OpenRLHF

14.8

Adjusted Score

14.8

Raw Score

100%

Time Factor

2026-07-14

Last Push

9.8K

Stars

Python

Language

16.7K

Lines of Code

Files

191

Pattern Hits

2026-07-14

Scan Date

0.00

HC Hit Rate

What These Metrics Mean

Adjusted Score: Primary synthetic code indicator. Raw score normalised per 1,000 lines of code and multiplied by the temporal discount factor. This is the definitive comparative metric — use it to rank repositories by AI authorship density.
Raw Score: The unmodified sum of all severity-weighted, context-multiplied pattern match scores before temporal discounting. Reflects the absolute signal strength independent of when the repository was last active.
Time Factor: The temporal discount multiplier (0–100%) applied to the raw score. Repositories last updated before ChatGPT's launch (Nov 2022) receive a 5% factor. Full signal is only assigned to repositories active in the post-adoption era (Jan 2024+).
Pattern Hits: Total count of individual pattern matches across all files and categories. A high hit count with a low score may indicate a very large codebase with isolated AI snippets; a low count with a high score indicates dense, concentrated AI signatures.
HC Hit Rate: High+Critical pattern hits per file, averaged across the repository. This orthogonal signal catches repositories where a few files are densely packed with high-severity AI tells — a strong indicator even when the normalised score appears moderate due to codebase size.
Lines of Code / Files: Total lines and files analysed. The scanner examines 94 file extensions. These denominators are used to normalise the score, enabling fair comparison between repositories of vastly different sizes.

Score History

Longitudinal tracking requires multiple scan runs. Once this repository is re-scanned after new commits land, this chart will visualise how the synthetic code signal evolves over time — enabling you to detect whether AI authorship is growing, stabilising, or being actively corrected by human engineers.

No multi-scan history yet — run the scanner again to build trend data.

Severity Breakdown

Classifies detected patterns by their diagnostic confidence and structural impact. CRITICAL patterns (coefficient 10) represent definitive synthetic signatures — hallucinated imports, explicit LLM attribution metadata — virtually never produced by human authors. HIGH (5) indicates strong structural tells such as cross-file repetition or cross-linguistic idioms. MEDIUM (2) covers recognisable conversational padding and AI-specific vocabulary. LOW (1) captures subtle indicators like tautological comments and generic boilerplate that require density to carry independent signal.

CRITICAL 0HIGH 0MEDIUM 19LOW 172

Directory Score Breakdown

This horizontal bar chart decomposes the repository's raw synthetic code score by top-level directory, allowing you to pinpoint precisely which modules or components carry the highest AI authorship density. Directories with disproportionately high scores relative to their size warrant targeted manual review: concentrated AI signatures often trace back to mass-generated configuration layers, auto-ported test suites, LLM-scaffolded boilerplate classes, or entire subsystems authored under heavy copilot assistance. Use this view to prioritise your human code-review effort.

Pattern Findings

The scanner identified 191 distinct pattern matches across 13 syntactic categories. Each entry below represents a discrete location in the source code where the engine recorded a statistically significant AI authorship indicator. Expand any category row to inspect the individual file paths, line numbers, code snippets, and the lexical context (CODE, COMMENT, or STRING) in which each match was detected.

Reading the findings table: The Severity column indicates the diagnostic confidence level (CRITICAL / HIGH / MEDIUM / LOW). The Context column identifies whether the match occurred inside executable code, an inline comment, or a string literal — comment-context matches receive a ×1.5 weight because LLMs systematically over-annotate. The ⚡ bolt icon marks clustered matches: three or more patterns within a 10-line window, each receiving an additional ×1.5 density multiplier as dense clusters constitute far stronger evidence of synthetic authorship than isolated hits.

Hyper-Verbose Identifiers55 hits · 62 pts

Severity	File	Line	Snippet	Context
LOW	openrlhf/utils/vlm_utils.py	68	def process_prompt_with_images(	CODE
LOW	openrlhf/utils/seqlen_balancing.py	153	def get_seqlen_balanced_partitions(seqlen_list: List[int], k_partitions: int, equal_size: bool):	CODE
LOW	openrlhf/utils/seqlen_balancing.py	171	def _check_and_sort_partitions(partitions):	CODE
LOW	openrlhf/utils/seqlen_balancing.py	234	def get_minimum_num_micro_batch_size(total_lengths, max_tokens_per_gpu, ring_attn_size, ds_tensor_parallel_size):	CODE
LOW⚡	openrlhf/utils/math_utils.py	175	def _strip_properly_formatted_commas(expr: str) -> str:	CODE
LOW	openrlhf/utils/math_utils.py	200	def _inject_implicit_mixed_number(step: str) -> str:	CODE
LOW	openrlhf/utils/math_utils.py	278	def count_unknown_letters_in_expr(expr: str) -> int:	CODE
LOW	openrlhf/utils/loss_utils.py	54	def _optimizer_step_loss_norm(masks, dp_group, dp_size, gas):	CODE
LOW	openrlhf/utils/loss_utils.py	80	def iter_grad_accum_global_norm(data_iter, strategy, accumulated_gradient, mask_fn):	CODE
LOW	openrlhf/utils/distributed_util.py	1	def torch_dist_barrier_and_cuda_sync():	CODE
LOW	openrlhf/utils/distributed_util.py	13	def stateless_init_process_group(master_address, master_port, rank, world_size, device):	CODE
LOW	openrlhf/utils/deepspeed/deepspeed_utils.py	121	def get_optimizer_grouped_parameters(	CODE
LOW	openrlhf/models/ring_attn_utils.py	9	def patch_transformers_for_ring_flash_attn():	CODE
LOW	openrlhf/models/ring_attn_utils.py	32	def reset_ring_attn_position_ids(start, end, packed_seq_lens):	CODE
LOW	openrlhf/models/ring_attn_utils.py	72	def get_tensor_in_current_ring_attn_rank(tensors: list[torch.Tensor] \| torch.Tensor, ring_attn_group, pad_id):	CODE
LOW	openrlhf/models/model.py	21	def get_llm_for_sequence_regression(	CODE
LOW⚡	openrlhf/models/actor.py	298	def gradient_checkpointing_enable(self, gradient_checkpointing_kwargs={"use_reentrant": False}):	CODE
LOW⚡	openrlhf/models/actor.py	301	def gradient_checkpointing_disable(self):	CODE
LOW⚡	openrlhf/models/actor.py	304	def print_trainable_parameters(self):	CODE
LOW	openrlhf/trainer/sft_trainer.py	215	def save_logs_and_checkpoints(self, args, global_step, step_bar, logs_dict=None, client_states=None):	CODE
LOW	openrlhf/trainer/rm_trainer.py	214	def save_logs_and_checkpoints(self, args, global_step, step_bar, logs_dict={}, client_states={}):	CODE
LOW	openrlhf/trainer/ppo_trainer.py	197	def restore_best_checkpoint_state(self, checkpoint_states) -> None:	CODE
LOW	openrlhf/trainer/ppo_trainer.py	383	def save_logs_and_checkpoints(self, global_step: int, logs_dict=None, client_states=None) -> None:	CODE
LOW	openrlhf/trainer/dpo_trainer.py	223	def save_logs_and_checkpoints(self, args, global_step, step_bar, logs_dict=None, client_states=None):	CODE
LOW	openrlhf/trainer/ppo_utils/samples_generator.py	199	def _dispatch_prompts_to_vllm(	CODE
LOW	openrlhf/trainer/ppo_utils/samples_generator.py	247	def _process_response_into_experience(self, response, **generate_kwargs) -> Experience:	CODE
LOW	openrlhf/trainer/ppo_utils/experience_maker.py	238	def compute_advantages_and_returns(self, experiences: List[Experience]) -> List[Experience]:	CODE
LOW	openrlhf/trainer/ppo_utils/experience_maker.py	333	def get_advantages_and_returns(	CODE
LOW	openrlhf/trainer/ppo_utils/experience.py	256	def remove_padding_in_sequences(items: List[Experience]) -> List[Experience]:	CODE
LOW	openrlhf/trainer/ppo_utils/length_penalty.py	61	def apply_stop_properly_penalty(	CODE
LOW	openrlhf/trainer/ray/vllm_engine.py	179	def get_num_unfinished_requests(self) -> int:	CODE
LOW	openrlhf/trainer/ray/ppo_actor.py	494	def init_model_from_pretrained(self, strategy: DeepspeedStrategy, pretrain, max_steps=None, vllm_engines=None):	CODE
LOW	openrlhf/trainer/ray/ppo_critic.py	189	def init_model_from_pretrained(self, strategy: DeepspeedStrategy, pretrain, max_steps):	CODE
LOW	openrlhf/trainer/ray/utils.py	20	def ray_noset_visible_devices(env_vars=os.environ):	CODE
LOW	openrlhf/trainer/ray/launcher.py	60	def init_model_from_pretrained(self, args, *kwargs):	CODE
LOW	openrlhf/trainer/ray/launcher.py	106	def init_model_from_pretrained(self, strategy: DeepspeedStrategy, pretrain):	CODE
LOW	openrlhf/trainer/ray/launcher.py	159	def init_model_from_pretrained(self, strategy: DeepspeedStrategy, pretrain):	CODE
LOW	openrlhf/trainer/ray/launcher.py	292	def async_init_model_from_pretrained(	CODE
LOW⚡	tests/test_ray_env_vars.py	140	def test_tokenizers_parallelism_default(self):	CODE
LOW⚡	tests/test_ray_env_vars.py	146	def test_ray_zero_copy_default(self):	CODE
LOW⚡	tests/test_ray_env_vars.py	161	def test_nccl_debug_user_info(self):	CODE
LOW⚡	tests/test_ray_env_vars.py	165	def test_nccl_debug_user_trace(self):	CODE
LOW⚡	tests/test_ray_env_vars.py	169	def test_tokenizers_parallelism_user_false(self):	CODE
LOW⚡	tests/test_ray_env_vars.py	173	def test_ray_zero_copy_user_disabled(self):	CODE
LOW⚡	tests/test_ray_env_vars.py	177	def test_multiple_user_overrides(self):	CODE
LOW⚡	tests/test_ray_env_vars.py	198	def test_ray_init_not_called_when_initialized(self):	CODE
LOW	tests/test_loss_aggregation.py	47	def test_token_mean_aggregation_matches_verl_dp_average():	CODE
LOW	tests/test_loss_aggregation.py	61	def test_seq_mean_token_mean_aggregation_matches_verl_dp_average():	CODE
LOW	tests/test_loss_aggregation.py	86	def test_loss_batch_info_excludes_fully_masked_sequences():	CODE
LOW	tests/test_loss_aggregation.py	97	def test_policy_loss_token_mean_aggregation_matches_full_batch():	CODE
LOW	tests/test_loss_aggregation.py	127	def test_policy_kl_metric_uses_policy_ratio_when_vllm_correction_is_enabled():	CODE
LOW	tests/test_loss_aggregation.py	151	def test_grad_accum_global_norm_matches_global_token_mean_over_window():	CODE
LOW	tests/test_loss_aggregation.py	183	def test_grad_accum_global_norm_differs_from_per_microbatch_when_uneven():	CODE
LOW	tests/test_loss_aggregation.py	196	def test_policy_kl_metric_is_not_clamped():	CODE
LOW⚡	examples/python/vlm_multiturn_agent.py	42	def get_screenshot_from_environment():	CODE

Decorative Section Separators12 hits · 45 pts

Severity	File	Line	Snippet	Context
MEDIUM	tests/test_ray_env_vars.py	126	# ---------------------------------------------------------------------------	COMMENT
MEDIUM	tests/test_ray_env_vars.py	128	# ---------------------------------------------------------------------------	COMMENT
MEDIUM⚡	tests/test_ray_env_vars.py	153	# ---------------------------------------------------------------------------	COMMENT
MEDIUM⚡	tests/test_ray_env_vars.py	155	# ---------------------------------------------------------------------------	COMMENT
MEDIUM⚡	tests/test_ray_env_vars.py	190	# ---------------------------------------------------------------------------	COMMENT
MEDIUM⚡	tests/test_ray_env_vars.py	192	# ---------------------------------------------------------------------------	COMMENT
MEDIUM⚡	examples/python/vlm_multiturn_agent.py	37	# ──────────────────────────────────────────────────────────────────────	COMMENT
MEDIUM⚡	examples/python/vlm_multiturn_agent.py	39	# ──────────────────────────────────────────────────────────────────────	COMMENT
MEDIUM	examples/python/vlm_multiturn_agent.py	69	# ──────────────────────────────────────────────────────────────────────	COMMENT
MEDIUM	examples/python/vlm_multiturn_agent.py	71	# ──────────────────────────────────────────────────────────────────────	COMMENT
MEDIUM	examples/python/vlm_multiturn_agent.py	177	# ──────────────────────────────────────────────────────────────────────	COMMENT
MEDIUM	examples/python/vlm_multiturn_agent.py	179	# ──────────────────────────────────────────────────────────────────────	COMMENT

Unused Imports32 hits · 32 pts

Severity	File	Line	Context
LOW	openrlhf/datasets/__init__.py	1	CODE
LOW	openrlhf/datasets/__init__.py	2	CODE
LOW	openrlhf/datasets/__init__.py	3	CODE
LOW	openrlhf/utils/math_utils.py	6	CODE
LOW	openrlhf/utils/__init__.py	1	CODE
LOW	openrlhf/utils/__init__.py	1	CODE
LOW	openrlhf/utils/__init__.py	2	CODE
LOW	openrlhf/utils/__init__.py	2	CODE
LOW	openrlhf/utils/deepspeed/__init__.py	1	CODE
LOW	openrlhf/models/__init__.py	1	CODE
LOW	openrlhf/models/__init__.py	2	CODE
LOW	openrlhf/models/__init__.py	2	CODE
LOW	openrlhf/models/__init__.py	2	CODE
LOW	openrlhf/models/__init__.py	2	CODE
LOW	openrlhf/models/__init__.py	2	CODE
LOW	openrlhf/models/__init__.py	2	CODE
LOW	openrlhf/models/__init__.py	2	CODE
LOW	openrlhf/models/__init__.py	2	CODE
LOW	openrlhf/models/__init__.py	12	CODE
LOW	openrlhf/trainer/ppo_utils/__init__.py	1	CODE
LOW	openrlhf/trainer/ppo_utils/__init__.py	1	CODE
LOW	openrlhf/trainer/ppo_utils/__init__.py	1	CODE
LOW	openrlhf/trainer/ppo_utils/__init__.py	1	CODE
LOW	openrlhf/trainer/ppo_utils/__init__.py	2	CODE
LOW	openrlhf/trainer/ppo_utils/__init__.py	2	CODE
LOW	openrlhf/trainer/ppo_utils/__init__.py	3	CODE
LOW	openrlhf/trainer/ppo_utils/__init__.py	3	CODE
LOW	openrlhf/trainer/ppo_utils/__init__.py	3	CODE
LOW	openrlhf/trainer/ppo_utils/__init__.py	4	CODE
LOW	openrlhf/trainer/ray/__init__.py	2	CODE
LOW	openrlhf/trainer/ray/__init__.py	2	CODE
LOW	examples/python/agent_func_openai_server_executor.py	16	CODE

Deep Nesting30 hits · 30 pts

Severity	File	Line	Context
LOW	openrlhf/datasets/prompts_dataset.py	26	CODE
LOW	openrlhf/datasets/utils.py	10	CODE
LOW	openrlhf/utils/vlm_utils.py	30	CODE
LOW	openrlhf/utils/math_utils.py	25	CODE
LOW	openrlhf/utils/math_utils.py	285	CODE
LOW	openrlhf/utils/math_utils.py	372	CODE
LOW	openrlhf/utils/math_utils.py	26	CODE
LOW	openrlhf/utils/agent.py	202	CODE
LOW	openrlhf/utils/deepspeed/deepspeed.py	507	CODE
LOW	openrlhf/utils/deepspeed/deepspeed.py	537	CODE
LOW	openrlhf/utils/deepspeed/deepspeed.py	683	CODE
LOW	openrlhf/models/model.py	21	CODE
LOW	openrlhf/models/utils.py	10	CODE
LOW	openrlhf/models/actor.py	76	CODE
LOW	openrlhf/models/actor.py	221	CODE
LOW	openrlhf/trainer/sft_trainer.py	215	CODE
LOW	openrlhf/trainer/sft_trainer.py	246	CODE
LOW	openrlhf/trainer/ppo_trainer_async.py	100	CODE
LOW	openrlhf/trainer/rm_trainer.py	214	CODE
LOW	openrlhf/trainer/rm_trainer.py	245	CODE
LOW	openrlhf/trainer/dpo_trainer.py	223	CODE
LOW	openrlhf/trainer/dpo_trainer.py	254	CODE
LOW	openrlhf/trainer/ppo_utils/samples_generator.py	140	CODE
LOW	openrlhf/trainer/ppo_utils/experience_maker.py	238	CODE
LOW	openrlhf/trainer/ppo_utils/experience.py	184	CODE
LOW	openrlhf/trainer/ppo_utils/experience.py	219	CODE
LOW	openrlhf/trainer/ppo_utils/experience.py	123	CODE
LOW	openrlhf/trainer/ppo_utils/length_penalty.py	61	CODE
LOW	openrlhf/trainer/ray/ppo_actor.py	155	CODE
LOW	examples/python/agent_func_openai_server_executor.py	241	CODE

Excessive Try-Catch Wrapping16 hits · 17 pts

Severity	File	Line	Snippet	Context
LOW	openrlhf/datasets/utils.py	62	except Exception as e:	CODE
LOW	openrlhf/utils/vlm_utils.py	63	except Exception as e:	CODE
LOW	openrlhf/utils/math_utils.py	38	except Exception:	CODE
LOW	openrlhf/utils/math_utils.py	66	except Exception:	CODE
LOW	openrlhf/utils/math_utils.py	130	except Exception:	CODE
LOW⚡	openrlhf/utils/math_utils.py	160	except Exception:	CODE
LOW⚡	openrlhf/utils/math_utils.py	167	except Exception:	CODE
LOW	openrlhf/utils/math_utils.py	190	except Exception:	CODE
LOW	openrlhf/utils/math_utils.py	248	except Exception:	CODE
LOW	openrlhf/utils/math_utils.py	311	except Exception:	CODE
LOW	openrlhf/utils/math_utils.py	361	except Exception:	CODE
LOW	openrlhf/utils/utils.py	47	except Exception:	CODE
LOW	openrlhf/utils/agent.py	298	except Exception as e:	CODE
LOW	openrlhf/utils/agent.py	347	except Exception as e: # pragma: no cover - defensive	CODE
LOW	examples/python/agent_func_openai_server_executor.py	47	except Exception:	CODE
LOW	examples/python/agent_func_openai_server_executor.py	218	except Exception:	CODE

Over-Commented Block12 hits · 12 pts

Severity	File	Line	Snippet	Context
LOW	README_zh.md	481	--vllm.sync_backend nccl \	COMMENT
LOW	openrlhf/utils/seqlen_balancing.py	1	# Copied from https://github.com/volcengine/verl/blob/468adf22c43b744348051fccd7a5d830c6c3c36a/verl/utils/seqlen_balanci	COMMENT
LOW	openrlhf/utils/deepspeed/deepspeed.py	141	# ``optim`` defaults to the strategy-wide setting but can be overridden per-model	COMMENT
LOW	openrlhf/utils/deepspeed/deepspeed.py	301	# sets are silently dropped. ns_steps / nesterov are NOT configurable via	COMMENT
LOW	examples/scripts/train_nonrl_slurm.sh	1	#!/bin/bash	COMMENT
LOW	examples/scripts/train_vlm_math_hybrid_engine.sh	1	#!/bin/bash	COMMENT
LOW	examples/scripts/train_vlm_math_hybrid_engine.sh	81	--ds.enable_sleep	COMMENT
LOW	examples/scripts/train_vlm_math_hybrid_engine.sh	101	# Images in the dataset should be a list of paths/URLs.	COMMENT
LOW	examples/scripts/train_prorlv2_math_hybrid_engine.sh	1	#!/bin/bash	COMMENT
LOW	examples/scripts/train_prorlv2_math_hybrid_engine.sh	81		COMMENT
LOW	examples/scripts/train_prorlv2_math_hybrid_engine.sh	101	# Truncated sample rewards are scaled by this coefficient (0.0 = zero reward for truncated)	COMMENT
LOW	examples/scripts/train_ppo_ray_slurm.sh	1	#!/bin/bash	COMMENT

Self-Referential Comments4 hits · 12 pts

Severity	File	Line	Snippet	Context
MEDIUM	openrlhf/utils/agent.py	78	# Initialize tracking variables	COMMENT
MEDIUM	openrlhf/models/utils.py	112	# The following code is equivalent to:	COMMENT
MEDIUM	openrlhf/trainer/ray/vllm_engine.py	248	# Create a big placement group to ensure that all engines are packed	COMMENT
MEDIUM	openrlhf/trainer/ray/utils.py	5	# This function is used to get the bundle indices of a placement group	COMMENT

Modern Structural Boilerplate12 hits · 12 pts

Severity	File	Line	Snippet	Context
LOW	openrlhf/datasets/__init__.py	5	__all__ = ["PromptDataset", "RewardDataset", "SFTDataset"]	CODE
LOW	openrlhf/utils/vlm_utils.py	15	logger = logging.getLogger(__name__)	CODE
LOW	openrlhf/utils/__init__.py	4	__all__ = [	CODE
LOW	openrlhf/utils/distributed_sampler.py	9	__all__ = ["DistributedSampler"]	CODE
LOW	openrlhf/utils/distributed_sampler.py	138	def set_epoch(self, epoch: int, consumed_samples=0) -> None:	STRING
LOW	openrlhf/utils/deepspeed/__init__.py	3	__all__ = [	CODE
LOW	openrlhf/models/__init__.py	14	__all__ = [	CODE
LOW	openrlhf/models/utils.py	10	def set_z3_leaf_modules(model: nn.Module, detect_hybrid: bool = True) -> None:	CODE
LOW	openrlhf/trainer/ppo_utils/__init__.py	6	__all__ = [	CODE
LOW	openrlhf/trainer/ray/__init__.py	4	__all__ = [	CODE
LOW	examples/python/agent_func_gem_multiturn.py	13	logger = logging.getLogger(__name__)	CODE
LOW	examples/python/agent_func_openai_server_executor.py	33	logger = logging.getLogger(__name__)	CODE

Redundant / Tautological Comments7 hits · 10 pts

Severity	File	Line	Snippet	Context
LOW	openrlhf/utils/agent.py	246	# Check if response was truncated (hit max_tokens length limit)	COMMENT
LOW	openrlhf/trainer/ppo_utils/samples_generator.py	281	# Check if response was truncated (hit max_tokens limit, finish_reason == "length")	COMMENT
LOW	openrlhf/trainer/ray/launcher.py	330	# Check if all kwargs parameters are iterable	COMMENT
LOW	examples/python/agent_func.py	47	# Check if episode is done	COMMENT
LOW	examples/python/agent_func_gem_multiturn.py	113	# Check if episode is done	COMMENT
LOW	examples/scripts/ckpt_ds_zero_to_universal.sh	10	# Set CKPT_PATH to the first argument and shift it out so that "$@" contains the extra arguments.	COMMENT
LOW	examples/scripts/ckpt_ds_zero_to_universal.sh	20	# Check if the latest tag exists.	COMMENT

AI Structural Patterns6 hits · 6 pts

Severity	File	Line	Context
LOW	openrlhf/utils/deepspeed/deepspeed_utils.py	7	CODE
LOW	openrlhf/models/model.py	21	CODE
LOW	openrlhf/models/loss.py	121	CODE
LOW	openrlhf/models/actor.py	76	CODE
LOW	openrlhf/models/actor.py	221	CODE
LOW	openrlhf/trainer/ray/ppo_actor.py	39	CODE

Modern AI Meta-Vocabulary2 hits · 5 pts

Severity	File	Line	Snippet	Context
MEDIUM	README.md	41	- [🎯 Design Paradigm](#design-paradigm-agent-based-execution) - Unified agent-based execution pipeline	CODE
MEDIUM	openrlhf/trainer/ppo_utils/samples_generator.py	264	# Truncate everything to the configured context window.	COMMENT

AI Slop Vocabulary1 hit · 3 pts

Severity	File	Line	Snippet	Context
MEDIUM	examples/scripts/train_prorlv2_math_hybrid_engine.sh	13	# ProRL v2 achieves state-of-the-art performance among 1.5B reasoning models	COMMENT

Overly Generic Function Names2 hits · 2 pts

Severity	File	Line	Snippet	Context
LOW	openrlhf/datasets/sft_dataset.py	89	def process_data(self, data):	CODE
LOW	openrlhf/datasets/reward_dataset.py	101	def process_data(self, data):	CODE

Analysis Overview

What These Metrics Mean

Score History

Severity Breakdown

Directory Score Breakdown

Pattern Findings