{
"$schema": "https://inlang.com/schema/inlang-message-format",
"hello_world": "Hello, {name} from en!",
"navigation_ranking": "Ranking",
"navigation_details": "Details",
"navigation_visualizations": "Visualizations",
"navigation_methodology": "Methodology",
"brand_guard_leaderboard": "GuardRate Leaderboard",
"brand_home_aria_label": "GuardRate Leaderboard by HiveTrace",
"ranking_meta_title": "Ranking - HiveTrace Guard Leaderboard",
"ranking_badge": "Guardrail benchmark",
"ranking_updated": "updated {date}",
"ranking_title": "Ranking",
"ranking_description": "A ranking of open-source guardrail models.",
"ranking_filters": "Ranking Filters",
"ranking_table": "Models Table",
"ranking_scope_select_label": "data slice",
"ranking_scope_all": "All benchmarks",
"ranking_scope_all_description": "19 groups",
"ranking_scope_english": "English slice",
"ranking_scope_english_description": "without RU groups",
"ranking_scope_robust": "Robustness & injection",
"ranking_scope_robust_description": "5 groups",
"ranking_csv": "Export CSV",
"ranking_filter_bar_label": "Leaderboard filters",
"ranking_filter_language": "Language",
"ranking_filter_models": "Models",
"ranking_filter_group_benchmark": "Group / benchmark",
"ranking_filter_metric": "Metric",
"ranking_filter_slice": "Slice",
"ranking_filter_language_any": "Any",
"ranking_filter_language_en": "EN",
"ranking_filter_language_ru": "RU",
"ranking_filter_select_all": "Select all",
"ranking_filter_models_all": "Models: All",
"ranking_filter_models_llm_judge": "LLM judges",
"ranking_filter_models_classifier": "Classifiers",
"ranking_filter_models_selected": "Models: {count}",
"ranking_filter_groups_all": "Groups: All",
"ranking_filter_groups_current": "Current slice",
"ranking_filter_groups_robust": "Robustness",
"ranking_filter_groups_selected": "Groups: {count}",
"ranking_filter_slice_combined": "combined",
"ranking_filter_slice_safe": "safe",
"ranking_filter_slice_unsafe": "unsafe",
"ranking_filter_summary": "{count} models · slice: {scope}",
"ranking_models_label": "Models",
"ranking_models_hint": "open-source guardrails",
"ranking_groups_label": "Groups in slice",
"ranking_best_integral_label": "Best Integral",
"ranking_fastest_p95_label": "Fastest p95",
"ranking_unit_ms": "ms",
"ranking_table_title": "Leaderboard",
"ranking_average_integral": "{scope}: average Integral {value}.",
"ranking_rows": "{count} rows",
"ranking_column_model": "Model",
"ranking_column_integral": "Integral",
"ranking_column_min_score": "Min score",
"ranking_column_fpr": "Overall FPR",
"ranking_column_fnr": "Overall FNR",
"ranking_column_p95": "p95 ms",
"ranking_column_size": "Size",
"ranking_column_higher_better": "higher is better",
"ranking_column_lower_better": "lower is better",
"ranking_column_report": "Report",
"ranking_model_type_classifier": "classifier",
"ranking_model_type_llm_judge": "llm judge",
"ranking_eval_leak": "eval-leak",
"ranking_eval_partial": "partial",
"ranking_open_details_label": "Open details for {model}",
"ranking_details_overview_tab": "Overview",
"ranking_details_groups_tab": "By group",
"ranking_details_benchmarks_tab": "Benchmarks",
"ranking_details_meta_tab": "Meta",
"ranking_details_overall_f1": "Overall F1",
"ranking_details_p50_p95": "p50 / p95 ms",
"ranking_details_full_name": "Full name",
"ranking_details_family": "Family",
"ranking_details_parameters": "Parameters",
"ranking_details_guard_type": "Guard type",
"ranking_details_languages": "Languages",
"ranking_details_license": "License",
"ranking_details_run_date": "Run date",
"ranking_details_coverage": "Coverage",
"ranking_details_coverage_value": "{covered}/{total} groups",
"ranking_details_group": "Group",
"ranking_details_benchmark": "Benchmark",
"ranking_details_score": "Score",
"ranking_details_repository": "Report / repository",
"ranking_details_open_hf": "Open on HF",
"benchmark_description_s_eval": "The S-Eval family (IS2Lab): a set of base risks and generated jailbreak attacks derived from them. It checks whether a model recognizes malicious intent in both original and disguised requests.",
"benchmark_description_aegis": "An evaluation of harmful-content filtering with a focus on safety and commercial applicability: the data is close to real moderation scenarios.",
"benchmark_description_toxic_chat": "A dialogue-toxicity benchmark based on real user interactions with LLMs. The final label is formed as an OR between the toxicity and jailbreaking columns.",
"benchmark_description_polyguard": "A universal high-quality benchmark based on WildGuardMix translated into 17 languages. It tests toxicity, adversarial attacks, unsafe requests, and false positives (exaggerated safety).",
"benchmark_description_rtp_lx": "ToxicityPrompts/RTP-LX: evaluates toxicity in a multilingual setting with cultural nuances and hidden forms of harm (microaggressions, bias) that often slip past models.",
"benchmark_description_or_bench": "A set of truly toxic prompts plus a hard 1k subset rejected by at least 3 of the largest SOTA models. It evaluates the safety/helpfulness balance while minimizing over-refusal.",
"benchmark_description_xstest": "walledai/XSTest: diagnoses model over-caution (false refusals on safe requests with trigger words) using contrastive pairs of safe and unsafe prompts.",
"benchmark_description_strong_reject": "Measures robustness to unambiguously harmful requests blocked by all market leaders. Answering such a prompt is treated as a critical safety failure (false negative). Translations were done by native speakers.",
"benchmark_description_beaver_tails": "Checks whether the model remains safe and appropriate by evaluating robustness to harmful-content generation on labeled prompt-response pairs and the balance between safety and helpfulness.",
"benchmark_description_harm_bench": "Human-labeled prompt-response pairs plus request subsets. It tests robustness to a broad range of adversarial attacks: AutoDan, GCG, Human, Few-shot, PAIR, Persona.",
"benchmark_description_multi_jail": "Evaluates whether the model correctly classifies human-translated real jailbreak requests across languages, including low-resource ones (en, zh, it, vi, ar, ko, th, bn, sw, jv).",
"benchmark_description_simple_safety_tests": "Bertievidgen/SimpleSafetyTests: a fast smoke test / sanity check made of concentrated harmful requests for quickly finding critical safety vulnerabilities.",
"benchmark_description_csrt": "walledai/CSRT: evaluates robustness to harmful requests disguised by mixing several languages in one phrase. It is built on top of MultiJail.",
"benchmark_description_aya_red_teaming": "CohereLabs/aya_redteaming: robustness to targeted attacks and correct handling of harmful requests across cultural contexts, with harm split into global and local categories.",
"benchmark_description_xsafety": "The first multilingual safety benchmark: 14 risk types across 10 languages. It tests robustness to unsafe requests and complex jailbreak attacks (role-playing, prompt leakage).",
"benchmark_description_openai_moderation": "mmathys/openai-moderation-api-evaluation: prevention of content that violates safety policy across a fixed set of categories (violence, hate, sexual content, self-harm).",
"benchmark_description_robustness_real": "Real user requests and responses: harmful, safe, and adversarial-safe splits. Aggregate FNR/FPR/F1 metrics assess overall robustness on native Russian data.",
"benchmark_description_robustness_robust": "Augmented (harder) versions of harmful and safe requests. Tests robustness to stronger attack variants and false positives on harder safe examples.",
"benchmark_description_prompt_injection": "A dataset for evaluating prompt-injection robustness: examples of attacks aimed at taking control of the model, in Russian and English.",
"benchmark_tests_s_eval": "Recognizing unsafe intent in clean requests (base risk) and robustness to 10 adversarial attack methods such as role-playing and forced-answer starts (attack set).",
"benchmark_tests_aegis": "Detecting harmful content in prompts and responses on realistic commercial data.",
"benchmark_tests_toxic_chat": "Classifying toxicity and jailbreak behavior in user inputs and model responses from real dialogue logs.",
"benchmark_tests_polyguard": "Detection robustness on English and Russian request/response slices, including over-refusal.",
"benchmark_tests_rtp_lx": "Detecting non-trivial cross-cultural toxicity in English and Russian requests and responses.",
"benchmark_tests_or_bench": "The ability to separate real harm from ambiguous requests without drifting into over-refusal.",
"benchmark_tests_xstest": "The decision boundary: false refusals (FPR) on safe requests versus missed real harm.",
"benchmark_tests_strong_reject": "Strict detection of explicit harm across five languages (EN, RU, UKR, BE, UZ); the priority is minimizing missed harmful content.",
"benchmark_tests_beaver_tails": "Detecting harm in model responses and identifying cases of excessive caution.",
"benchmark_tests_harm_bench": "Recognizing LLM-generated harm under attack and robustness to context in requests (standard/contextual) and responses.",
"benchmark_tests_multi_jail": "Multilingual detection of jailbreak requests in the English slice.",
"benchmark_tests_simple_safety_tests": "Baseline detection of explicit harm as a quick model sanity check.",
"benchmark_tests_csrt": "Robustness to code-switching attacks (multiple languages mixed in one prompt).",
"benchmark_tests_aya_red_teaming": "Adversarial robustness in Russian and English slices with culturally specific harms.",
"benchmark_tests_xsafety": "Safety degradation on non-English requests and robustness to adversarial scenarios in the English slice.",
"benchmark_tests_openai_moderation": "Conformance to a fixed moderation taxonomy based on safety policy.",
"benchmark_tests_robustness_real": "Recognizing real attacks while keeping FPR low on safe and adversarial-safe requests (request and response modes).",
"benchmark_tests_robustness_robust": "Robustness to augmented attacks: FNR on harm and FPR on safe splits of Russian data.",
"benchmark_tests_prompt_injection": "Detecting prompt-injection attacks across two language slices.",
"methodology_meta_title": "Methodology - HiveTrace Guard Leaderboard",
"methodology_title": "Methodology",
"methodology_description": "Dataset → group → integral score. FPR and FNR stay separate until the final aggregation step.",
"methodology_meaning_label": "What it means:",
"methodology_aggregation_title": "Three aggregation levels",
"methodology_dataset_score_tag": "dataset score",
"methodology_dataset_score_comment": "harmonic mean of (1−FPR, 1−FNR)",
"methodology_dataset_score_description": "Errors are normalized into quality: 1−FPR is the share of safe content correctly passed, and 1−FNR is the share of harmful content correctly caught. Then we take the harmonic mean, so the weaker side pulls the score down.",
"methodology_group_score_tag": "group score",
"methodology_group_score_comment": "harmonic mean of N dataset scores",
"methodology_group_score_description": "A group is a family of related datasets, for example jailbreak tests. We use the harmonic mean again: the weakest dataset defines the group score, so easy sets cannot hide a poor result inside the same topic.",
"methodology_integral_score_tag": "final score",
"methodology_integral_score_comment": "geometric mean of M groups (S_group > 0)",
"methodology_integral_log_comment": "log form for numerical stability",
"methodology_integral_score_description": "Across groups we use the geometric mean. It penalizes imbalance more strongly: one failing group pulls the product down. A balanced profile is valued above one isolated star result with a gap elsewhere.",
"methodology_combined_dataset_title": "Combined dataset",
"methodology_harm_split_tag": "harm split",
"methodology_harm_split_description": "The share of harmful requests incorrectly marked safe. TP are caught attacks, FN are missed attacks.",
"methodology_safe_split_tag": "safe split",
"methodology_safe_split_description": "The share of safe requests incorrectly marked harmful. TN are correctly passed safe items, FP are false alarms.",
"methodology_combined_dataset_note": "If a shared set is marked as combined, its metrics are taken directly from the full splits; otherwise metrics are computed in the standard way through the arithmetic mean of individual dataset metrics.",
"methodology_scatter_title": "Scatter: FPR vs FNR",
"methodology_scatter_x_axis": "Axis X",
"methodology_scatter_x_description": "Arithmetic mean of FPR across datasets in the group.",
"methodology_scatter_y_axis": "Axis Y",
"methodology_scatter_y_description": "Arithmetic mean of FNR across datasets in the group.",
"methodology_scatter_point_title": "Point = model",
"methodology_scatter_point_description": "The position shows the average balance between false blocks on the X axis and missed threats on the Y axis across all benchmarks. Closer to the origin is better on both axes.",
"methodology_specifics_title": "Calculation specifics",
"methodology_single_class_title": "Single-class datasets",
"methodology_single_class_description": "F1 is computed for all benchmarks by default. If a dataset contains labels from only one class: all_unsafe gets FNR, recall; all_safe respectively gets FPR, TNR.",
"methodology_caveat_title": "Caveat",
"methodology_caveat_description": "Metrics are point estimates. Confidence intervals and sample sizes (n) for tests are not published in the current dataset, so small ranking differences may be statistically insignificant.",
"ranking_report_link_label": "Open {model}",
"details": "Details",
"details_description": "Metrics for benchmarks and models.",
"benchmarks": "Benchmarks",
"benchmarks_description": "Summary matrix: model × test",
"benchmarks_filters": "Benchmark Filters",
"models": "Models",
"details_filter_all": "All",
"details_filter_type": "Type",
"details_filter_sort": "Sort",
"details_filter_search": "Search",
"details_filter_model": "Model",
"details_metric_recall": "Recall",
"details_metric_precision": "Precision",
"details_metric_accuracy": "Accuracy",
"details_benchmark_search_placeholder": "Benchmark name...",
"details_model_search_placeholder": "Search by name...",
"details_model_count": "{shown} of {total} models",
"details_error_rate": "Error rate",
"details_matrix_summary": "Summary: {metric} across all tests - {models} models × {tests} tests",
"details_matrix_model_sort": "Model · Integral",
"details_matrix_no_data": "no data",
"details_matrix_empty": "Nothing found. Reset search or enable groups.",
"details_dataset_matrix_unavailable_title": "Dataset matrix is temporarily unavailable",
"details_dataset_matrix_unavailable_description": "The current data includes dataset-level metrics for {covered} of {total} datasets. Group-level cards, rankings, and model profiles below use real data; the dataset matrix will be enabled when coverage is complete.",
"details_families_summary": "{families} families · {datasets} datasets",
"details_models_summary": "{families} families · {models} models",
"details_benchmark_families_title": "Benchmark families",
"details_benchmark_families_description": "Open a card for test details and model rankings.",
"details_family_stat_families": "families",
"details_family_stat_datasets": "datasets",
"details_family_stat_languages": "languages",
"details_family_stat_models": "models evaluated",
"details_family_filters": "Benchmark family filters",
"details_family_language_all": "All languages",
"details_family_empty": "Nothing found. Change filters or reset the search.",
"details_reset": "Reset",
"details_benchmark_sheet_what_tests": "What it tests",
"details_benchmark_sheet_datasets": "Datasets",
"details_benchmark_sheet_ranking": "Model ranking on this benchmark",
"details_benchmark_sheet_rank": "Rank",
"details_benchmark_sheet_open": "Open benchmark",
"details_benchmark_leader_metric": "LEADER · {metric}",
"details_model_sheet_open": "Open model",
"details_model_sheet_overview": "Overview",
"details_model_sheet_groups": "Groups",
"details_model_sheet_weak": "Benchmarks",
"details_model_sheet_meta": "Meta",
"details_model_sheet_min_group": "Min group",
"details_model_sheet_latency": "Latency p50/p95/p99",
"details_model_sheet_radar": "Group Radar",
"details_model_sheet_profile": "Group profile",
"details_model_sheet_profile_hint": "Group score (sorted, higher is better)",
"details_model_sheet_score_legend": "group score (higher is better)",
"details_model_sheet_groups_with_data": "{count} groups with data",
"details_model_sheet_weak_title": "Weakest benchmark groups",
"details_model_sheet_weak_hint": "5 groups with the lowest score. Higher is better for Recall/Precision and score/F1.",
"details_model_sheet_all_benchmarks": "Go to all benchmarks",
"details_model_sheet_full_name": "Full name",
"details_model_sheet_family": "Family",
"details_model_sheet_params": "Parameters",
"details_model_sheet_type": "Guardrail type",
"details_model_sheet_languages": "Languages",
"details_model_sheet_status": "Status",
"details_model_sheet_run_date": "Run date",
"details_model_sheet_links": "Links",
"details_model_sheet_hf_card": "HF card",
"details_dataset_count": "{count} datasets",
"details_no_data": "no data",
"details_license_label": "license",
"details_groups_label": "groups",
"details_groups_title": "Groups",
"details_group_other": "Other",
"details_select_all": "All",
"details_clear": "Clear",
"details_export": "Export",
"tools_meta_title": "Visualizations - HiveTrace Guard Leaderboard",
"tools_meta_description": "A visualization workspace for researchers: radar, FPR×FNR scatter, FNR heatmap, grouped bars, and latency Pareto.",
"tools_description": "The filters below control all charts in sync; “Pin” locks a model for comparison.",
"tools_controls_aria": "Visualization control panel",
"tools_models_title": "Models",
"tools_groups_title": "Groups / benchmarks",
"tools_language_label": "Language",
"tools_language_all": "All languages",
"tools_pin": "Pin",
"tools_unpin": "unpin",
"tools_pinned": "pinned",
"tools_radar_scale": "Radar scale",
"tools_radar_group_score": "group score",
"tools_mean": "mean",
"tools_group_mean_summary": "average across {groups} groups",
"tools_real_score": "score real",
"tools_robust_score": "score robust",
"tools_selected_summary": "{models} models · {groups} groups",
"tools_pinned_for_comparison": "Pinned for comparison",
"tools_unpin_model": "Unpin {model}",
"tools_no_pinned": "Nothing pinned, press Pin in the Models list.",
"tools_radar_title": "Group radar",
"tools_radar_hint": "scores of selected models by group · higher = better",
"tools_radar_aria": "Group radar for selected models",
"tools_radar_empty": "Need ≥3 groups with data for the radar.",
"tools_scatter_hint": "x = FPR, y = FNR · lower = better · ★ ideal (0,0)",
"tools_scatter_aria": "FPR versus FNR scatter",
"tools_scatter_ideal": "★ Ideal (0,0) - zero errors",
"tools_heatmap_scale_min": "min {value}",
"tools_heatmap_scale_max": "max {value}",
"tools_heatmap_header": "{metric} · group / model",
"tools_heatmap_empty": "Select at least one model and one group to show the heatmap.",
"tools_grouped_bars_title": "Grouped bars",
"tools_grouped_bars_hint": "Recall / Precision / F1 by model · averages across selected groups",
"tools_grouped_bars_aria": "Grouped Recall Precision F1 bars by model",
"tools_lower_better": "lower = better",
"tools_higher_better": "higher = better",
"tools_pareto_title": "Pareto: quality × latency",
"tools_pareto_hint": "x = p95 latency (ms, lower = better), y = integral (higher = better)",
"tools_pareto_aria": "Pareto quality versus p95 latency",
"tools_pareto_x_axis": "p95 ms ↓",
"tools_pareto_y_axis": "Integral ↑",
"tools_pareto_click_pin_hint": "hover for details · dashed line = Pareto frontier",
"tools_pareto_pin_model": "Pin {model} label",
"tools_pareto_unpin_model": "Unpin {model} label",
"tools_latency_title": "Latency performance",
"tools_latency_hint": "p50 / p95 / p99 (ms) · error rate",
"tools_latency_note": "Single-request latency percentiles.",
"tools_robustness_title": "Robustness real → robust",
"tools_robustness_hint": "degradation under augmentation - who breaks on harder attacks",
"tools_robustness_note": "Delta = robust - real. Delta score < 0: the model is not robust to obfuscations, > 0: robust. Delta FNR > 0: it misses harmful obfuscated messages more often (worse), < 0: it catches them more often (better). Delta FPR > 0: it blocks safe obfuscated messages more often (worse), < 0: it lets them pass more often (better).",
"tools_footer": "HiveTrace Guardrail Leaderboard",
"bucket_data_unavailable_title": "Data is temporarily unavailable",
"bucket_data_unavailable_description": "We could not load data for this page. Try refreshing the page again in a moment.",
"bucket_data_stale_title": "Showing saved data",
"bucket_data_stale_description": "The latest data refresh failed, so this page is using the last successfully loaded version.",
"not_found": "Page not found",
"the_page_youre_looking_for_doesnt_exist": "The page you’re looking for doesn’t exist.",
"go_back": "Go back",
"to_main_page": "Go to home page",
"internal_server_error": "Internal server error",
"sorry_something_went_wrong_on_our_end_we_are_currently_trying_to_fix_the_problem": "Sorry, something went wrong on our end. We’re working to fix the problem.",
"reload_page": "Reload page",
"snapshot_created": "snapshot: {date}"
}