jmrenouard/microllm-guardrail-int8
Text Classification • Updated • 32 • 2
Dataset Preview
The full dataset viewer is not available (click to read why). Only showing a preview of the rows.
The dataset generation failed
Error code: DatasetGenerationError
Exception: CastError
Message: Couldn't cast
id: string
domain: string
message: string
decision_type: string
instructions: string
rules: string
expected_result: string
ambiguous_edge: bool
expected_rubric: string
performance_sla: struct<p50_latency_ms: double, p95_latency_ms: double, p99_latency_ms: double, throughput_decisions_ (... 47 chars omitted)
child 0, p50_latency_ms: double
child 1, p95_latency_ms: double
child 2, p99_latency_ms: double
child 3, throughput_decisions_per_sec: double
child 4, int8_quantization_mse: double
total_decisions_evaluated: int64
quality_gates: struct<G1_jevbench_core_accuracy_ge_95pct: bool, G2_decision_pivot_pair_accuracy_ge_90pct: bool, G3_ (... 249 chars omitted)
child 0, G1_jevbench_core_accuracy_ge_95pct: bool
child 1, G2_decision_pivot_pair_accuracy_ge_90pct: bool
child 2, G3_anyjev_l0_zero_order_flip: bool
child 3, G4_guardrail_f1_ge_94pct: bool
child 4, G5_calibration_ece_and_brier_improved: bool
child 5, G6_jevcompat_contract_100pct_and_zero_5xx: bool
child 6, G7_multilingual_and_xai_ge_95pct: bool
child 7, G8_performance_sla_p50_le_10ms_and_int8_mse: bool
total_test_cases: int64
all_gates_passed: bool
tracks: struct<jevbench_core: struct<total: int64, gear1_accuracy: double, gear2_accuracy: double, cascade_a (... 2714 chars omitted)
child 0, jevbench_core: struct<total: int64, gear1_accuracy: double, gear2_accuracy: double, cascade_accuracy: double, by_pr (... 636 chars omitted)
child 0, total: int64
child 1, gear1_accuracy: double
...
p_5xx_count: int64
child 6, multilingual_and_xai_fidelity: struct<total_cases: int64, passed_cases: int64, pass_rate: double, xai_normalization_rate: double, b (... 195 chars omitted)
child 0, total_cases: int64
child 1, passed_cases: int64
child 2, pass_rate: double
child 3, xai_normalization_rate: double
child 4, by_language: struct<EN: struct<total: int64, accuracy: double>, FR: struct<total: int64, accuracy: double>, ES: s (... 82 chars omitted)
child 0, EN: struct<total: int64, accuracy: double>
child 0, total: int64
child 1, accuracy: double
child 1, FR: struct<total: int64, accuracy: double>
child 0, total: int64
child 1, accuracy: double
child 2, ES: struct<total: int64, accuracy: double>
child 0, total: int64
child 1, accuracy: double
child 3, DE: struct<total: int64, accuracy: double>
child 0, total: int64
child 1, accuracy: double
child 7, performance_sla_and_batch_stress: struct<total_cases: int64, passed_cases: int64, pass_rate: double, embedding_cache_hits: int64, int8 (... 33 chars omitted)
child 0, total_cases: int64
child 1, passed_cases: int64
child 2, pass_rate: double
child 3, embedding_cache_hits: int64
child 4, int8_mean_reconstruction_mse: double
suite_name: string
failed_gates: list<item: null>
child 0, item: null
total_elapsed_ms: double
to
{'suite_name': Value('string'), 'total_test_cases': Value('int64'), 'total_decisions_evaluated': Value('int64'), 'total_elapsed_ms': Value('float64'), 'all_gates_passed': Value('bool'), 'quality_gates': {'G1_jevbench_core_accuracy_ge_95pct': Value('bool'), 'G2_decision_pivot_pair_accuracy_ge_90pct': Value('bool'), 'G3_anyjev_l0_zero_order_flip': Value('bool'), 'G4_guardrail_f1_ge_94pct': Value('bool'), 'G5_calibration_ece_and_brier_improved': Value('bool'), 'G6_jevcompat_contract_100pct_and_zero_5xx': Value('bool'), 'G7_multilingual_and_xai_ge_95pct': Value('bool'), 'G8_performance_sla_p50_le_10ms_and_int8_mse': Value('bool')}, 'failed_gates': List(Value('null')), 'performance_sla': {'p50_latency_ms': Value('float64'), 'p95_latency_ms': Value('float64'), 'p99_latency_ms': Value('float64'), 'throughput_decisions_per_sec': Value('float64'), 'int8_quantization_mse': Value('float64')}, 'tracks': {'jevbench_core': {'total': Value('int64'), 'gear1_accuracy': Value('float64'), 'gear2_accuracy': Value('float64'), 'cascade_accuracy': Value('float64'), 'by_primitive': {'noul': {'total': Value('int64'), 'correct': Value('int64'), 'gear1_correct': Value('int64'), 'gear2_correct': Value('int64'), 'accuracy': Value('float64'), 'gear1_accuracy': Value('float64'), 'gear2_accuracy': Value('float64')}, 'choice': {'total': Value('int64'), 'correct': Value('int64'), 'gear1_correct': Value('int64'), 'gear2_correct': Value('int64'), 'accuracy': Value('float64'), 'gear1_accuracy': Value('float64'),
...
: Value('int64'), 'interception_or_pass_rate': Value('float64')}}}, 'calibration_rlcd2': {'total_items': Value('int64'), 'fitted_temperature_k4': Value('float64'), 'raw_brier': Value('float64'), 'calibrated_brier': Value('float64'), 'rlcd2_brier': Value('float64'), 'raw_ece': Value('float64'), 'calibrated_ece': Value('float64'), 'rlcd2_ece': Value('float64'), 'rlcd2_accuracy': Value('float64'), 'aurc': Value('float64'), 'selective_cov_at_5pct_error': Value('float64'), 'conformal_q_hat_90': Value('float64'), 'conformal_empirical_coverage': Value('float64'), 'clopper_pearson_certified_coverage': Value('float64'), 'proposition1_identity_verified': Value('bool')}, 'jevcompat_contract_and_boundaries': {'total_cases': Value('int64'), 'passed_cases': Value('int64'), 'pass_rate': Value('float64'), 'http_5xx_count': Value('int64')}, 'multilingual_and_xai_fidelity': {'total_cases': Value('int64'), 'passed_cases': Value('int64'), 'pass_rate': Value('float64'), 'xai_normalization_rate': Value('float64'), 'by_language': {'EN': {'total': Value('int64'), 'accuracy': Value('float64')}, 'FR': {'total': Value('int64'), 'accuracy': Value('float64')}, 'ES': {'total': Value('int64'), 'accuracy': Value('float64')}, 'DE': {'total': Value('int64'), 'accuracy': Value('float64')}}}, 'performance_sla_and_batch_stress': {'total_cases': Value('int64'), 'passed_cases': Value('int64'), 'pass_rate': Value('float64'), 'embedding_cache_hits': Value('int64'), 'int8_mean_reconstruction_mse': Value('float64')}}}
because column names don't match
Traceback: Traceback (most recent call last):
File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1827, in _prepare_split_single
for key, table in generator:
^^^^^^^^^
File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 613, in wrapped
for item in generator(*args, **kwargs):
~~~~~~~~~^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 343, in _generate_tables
self._cast_table(pa_table, json_field_paths=json_field_paths),
~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 132, in _cast_table
pa_table = table_cast(pa_table, self.info.features.arrow_schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2378, in table_cast
return cast_table_to_schema(table, schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2306, in cast_table_to_schema
raise CastError(
...<3 lines>...
)
datasets.table.CastError: Couldn't cast
id: string
domain: string
message: string
decision_type: string
instructions: string
rules: string
expected_result: string
ambiguous_edge: bool
expected_rubric: string
performance_sla: struct<p50_latency_ms: double, p95_latency_ms: double, p99_latency_ms: double, throughput_decisions_ (... 47 chars omitted)
child 0, p50_latency_ms: double
child 1, p95_latency_ms: double
child 2, p99_latency_ms: double
child 3, throughput_decisions_per_sec: double
child 4, int8_quantization_mse: double
total_decisions_evaluated: int64
quality_gates: struct<G1_jevbench_core_accuracy_ge_95pct: bool, G2_decision_pivot_pair_accuracy_ge_90pct: bool, G3_ (... 249 chars omitted)
child 0, G1_jevbench_core_accuracy_ge_95pct: bool
child 1, G2_decision_pivot_pair_accuracy_ge_90pct: bool
child 2, G3_anyjev_l0_zero_order_flip: bool
child 3, G4_guardrail_f1_ge_94pct: bool
child 4, G5_calibration_ece_and_brier_improved: bool
child 5, G6_jevcompat_contract_100pct_and_zero_5xx: bool
child 6, G7_multilingual_and_xai_ge_95pct: bool
child 7, G8_performance_sla_p50_le_10ms_and_int8_mse: bool
total_test_cases: int64
all_gates_passed: bool
tracks: struct<jevbench_core: struct<total: int64, gear1_accuracy: double, gear2_accuracy: double, cascade_a (... 2714 chars omitted)
child 0, jevbench_core: struct<total: int64, gear1_accuracy: double, gear2_accuracy: double, cascade_accuracy: double, by_pr (... 636 chars omitted)
child 0, total: int64
child 1, gear1_accuracy: double
...
p_5xx_count: int64
child 6, multilingual_and_xai_fidelity: struct<total_cases: int64, passed_cases: int64, pass_rate: double, xai_normalization_rate: double, b (... 195 chars omitted)
child 0, total_cases: int64
child 1, passed_cases: int64
child 2, pass_rate: double
child 3, xai_normalization_rate: double
child 4, by_language: struct<EN: struct<total: int64, accuracy: double>, FR: struct<total: int64, accuracy: double>, ES: s (... 82 chars omitted)
child 0, EN: struct<total: int64, accuracy: double>
child 0, total: int64
child 1, accuracy: double
child 1, FR: struct<total: int64, accuracy: double>
child 0, total: int64
child 1, accuracy: double
child 2, ES: struct<total: int64, accuracy: double>
child 0, total: int64
child 1, accuracy: double
child 3, DE: struct<total: int64, accuracy: double>
child 0, total: int64
child 1, accuracy: double
child 7, performance_sla_and_batch_stress: struct<total_cases: int64, passed_cases: int64, pass_rate: double, embedding_cache_hits: int64, int8 (... 33 chars omitted)
child 0, total_cases: int64
child 1, passed_cases: int64
child 2, pass_rate: double
child 3, embedding_cache_hits: int64
child 4, int8_mean_reconstruction_mse: double
suite_name: string
failed_gates: list<item: null>
child 0, item: null
total_elapsed_ms: double
to
{'suite_name': Value('string'), 'total_test_cases': Value('int64'), 'total_decisions_evaluated': Value('int64'), 'total_elapsed_ms': Value('float64'), 'all_gates_passed': Value('bool'), 'quality_gates': {'G1_jevbench_core_accuracy_ge_95pct': Value('bool'), 'G2_decision_pivot_pair_accuracy_ge_90pct': Value('bool'), 'G3_anyjev_l0_zero_order_flip': Value('bool'), 'G4_guardrail_f1_ge_94pct': Value('bool'), 'G5_calibration_ece_and_brier_improved': Value('bool'), 'G6_jevcompat_contract_100pct_and_zero_5xx': Value('bool'), 'G7_multilingual_and_xai_ge_95pct': Value('bool'), 'G8_performance_sla_p50_le_10ms_and_int8_mse': Value('bool')}, 'failed_gates': List(Value('null')), 'performance_sla': {'p50_latency_ms': Value('float64'), 'p95_latency_ms': Value('float64'), 'p99_latency_ms': Value('float64'), 'throughput_decisions_per_sec': Value('float64'), 'int8_quantization_mse': Value('float64')}, 'tracks': {'jevbench_core': {'total': Value('int64'), 'gear1_accuracy': Value('float64'), 'gear2_accuracy': Value('float64'), 'cascade_accuracy': Value('float64'), 'by_primitive': {'noul': {'total': Value('int64'), 'correct': Value('int64'), 'gear1_correct': Value('int64'), 'gear2_correct': Value('int64'), 'accuracy': Value('float64'), 'gear1_accuracy': Value('float64'), 'gear2_accuracy': Value('float64')}, 'choice': {'total': Value('int64'), 'correct': Value('int64'), 'gear1_correct': Value('int64'), 'gear2_correct': Value('int64'), 'accuracy': Value('float64'), 'gear1_accuracy': Value('float64'),
...
: Value('int64'), 'interception_or_pass_rate': Value('float64')}}}, 'calibration_rlcd2': {'total_items': Value('int64'), 'fitted_temperature_k4': Value('float64'), 'raw_brier': Value('float64'), 'calibrated_brier': Value('float64'), 'rlcd2_brier': Value('float64'), 'raw_ece': Value('float64'), 'calibrated_ece': Value('float64'), 'rlcd2_ece': Value('float64'), 'rlcd2_accuracy': Value('float64'), 'aurc': Value('float64'), 'selective_cov_at_5pct_error': Value('float64'), 'conformal_q_hat_90': Value('float64'), 'conformal_empirical_coverage': Value('float64'), 'clopper_pearson_certified_coverage': Value('float64'), 'proposition1_identity_verified': Value('bool')}, 'jevcompat_contract_and_boundaries': {'total_cases': Value('int64'), 'passed_cases': Value('int64'), 'pass_rate': Value('float64'), 'http_5xx_count': Value('int64')}, 'multilingual_and_xai_fidelity': {'total_cases': Value('int64'), 'passed_cases': Value('int64'), 'pass_rate': Value('float64'), 'xai_normalization_rate': Value('float64'), 'by_language': {'EN': {'total': Value('int64'), 'accuracy': Value('float64')}, 'FR': {'total': Value('int64'), 'accuracy': Value('float64')}, 'ES': {'total': Value('int64'), 'accuracy': Value('float64')}, 'DE': {'total': Value('int64'), 'accuracy': Value('float64')}}}, 'performance_sla_and_batch_stress': {'total_cases': Value('int64'), 'passed_cases': Value('int64'), 'pass_rate': Value('float64'), 'embedding_cache_hits': Value('int64'), 'int8_mean_reconstruction_mse': Value('float64')}}}
because column names don't match
The above exception was the direct cause of the following exception:
Traceback (most recent call last):
File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 1369, in compute_config_parquet_and_info_response
parquet_operations, partial, estimated_dataset_info = stream_convert_to_parquet(
~~~~~~~~~~~~~~~~~~~~~~~~~^
builder, max_dataset_size_bytes=max_dataset_size_bytes
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
)
^
File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 948, in stream_convert_to_parquet
builder._prepare_split(split_generator=splits_generators[split], file_format="parquet")
~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1694, in _prepare_split
for job_id, done, content in self._prepare_split_single(
~~~~~~~~~~~~~~~~~~~~~~~~~~^
gen_kwargs=gen_kwargs, job_id=job_id, **_prepare_split_args
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
):
^
File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1879, in _prepare_split_single
raise DatasetGenerationError("An error occurred while generating the dataset") from e
datasets.exceptions.DatasetGenerationError: An error occurred while generating the datasetNeed help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
suite_name string | total_test_cases int64 | total_decisions_evaluated int64 | total_elapsed_ms float64 | all_gates_passed bool | quality_gates dict | failed_gates list | performance_sla dict | tracks dict |
|---|---|---|---|---|---|---|---|---|
MicroLLM-Quality-Validation-1000 | 1,000 | 2,144 | 5,788.17 | true | {
"G1_jevbench_core_accuracy_ge_95pct": true,
"G2_decision_pivot_pair_accuracy_ge_90pct": true,
"G3_anyjev_l0_zero_order_flip": true,
"G4_guardrail_f1_ge_94pct": true,
"G5_calibration_ece_and_brier_improved": true,
"G6_jevcompat_contract_100pct_and_zero_5xx": true,
"G7_multilingual_and_xai_ge_95pct": true... | [] | {
"p50_latency_ms": 0.433,
"p95_latency_ms": 1.427,
"p99_latency_ms": 2.157,
"throughput_decisions_per_sec": 370.4,
"int8_quantization_mse": 0.0000041
} | {
"jevbench_core": {
"total": 160,
"gear1_accuracy": 0.9625,
"gear2_accuracy": 0.9875,
"cascade_accuracy": 0.9875,
"by_primitive": {
"noul": {
"total": 40,
"correct": 38,
"gear1_correct": 34,
"gear2_correct": 38,
"accuracy": 0.95,
"gear1_accura... |
MicroLLM-Quality-Validation-750 | 750 | 1,433 | 4,968.25 | true | {
"G1_jevbench_core_accuracy_ge_95pct": true,
"G2_decision_pivot_pair_accuracy_ge_90pct": true,
"G3_anyjev_l0_zero_order_flip": true,
"G4_guardrail_f1_ge_94pct": true,
"G5_calibration_ece_and_brier_improved": true,
"G6_jevcompat_contract_100pct_and_zero_5xx": true,
"G7_multilingual_and_xai_ge_95pct": true... | [] | {
"p50_latency_ms": 0.294,
"p95_latency_ms": 1.281,
"p99_latency_ms": 1.518,
"throughput_decisions_per_sec": 288.4,
"int8_quantization_mse": 0.0000041
} | {
"jevbench_core": {
"total": 160,
"gear1_accuracy": 0.9625,
"gear2_accuracy": 0.9875,
"cascade_accuracy": 0.9875,
"by_primitive": {
"noul": {
"total": 40,
"correct": 38,
"gear1_correct": 34,
"gear2_correct": 38,
"accuracy": 0.95,
"gear1_accura... |
⚡ MicroLLM — Master Benchmark Suite & Comprehensive Evaluation Analysis
🌐 Official GitHub Repository: https://github.com/jmrenouard/MicroLLM
🚀 Interactive Demo (Hugging Face Space): https://huggingface.co/spaces/jmrenouard/microllm-decision-studio
📜 JevCompat 0.1 Specification: 32/32 MUST protocol compliance (POST /v1/systemone)
This repository hosts all canonical evaluation telemetry, raw JSON/JSONL benchmarks, and deep-dive analysis reports for MicroLLM, a local probabilistic System-1 decision engine engineered for sub-millisecond CPU latency (< 0.4 ms, 0% GPU, < 85 MB RAM, $0 API Cost).
📊 Master Verification Scorecard (9 Suites · 128,497 Items)
All computations executed strictly on a single consumer laptop CPU core (AMD Ryzen AI 9 365, 0% GPU, < 85 MB RAM):
| Benchmark Suite | Total Items | Primary Accuracy | Order-Flip Rate | CPU Latency (p50) | Verification Status | Artifact File |
|---|---|---|---|---|---|---|
Structured Universal (11k) |
11,000 |
99.75% (100% Selective) |
0.000000 |
0.076–0.190 ms |
PASS (10,973/11,000) | test_cases_11000_results.json |
Unsampled External (16 Datasets) |
114,018 |
87.49% Selective (66.82% raw) |
0.000000 |
0.369–0.938 ms |
PASS (114,018 rows) | web_benchmarks_full_results.json |
Scientific Core (540) |
540 |
98.75% (100% Pair Acc) |
0.000000 |
0.266 ms |
PASS (540/540) | latest_results.json |
Quality SLA (1,000) |
1,000 |
98.75% Core |
0.000000 |
0.433 ms |
PASS (8/8 Gates) | quality_validation_1000_results.json |
Quality SLA (750) |
750 |
98.75% Core |
0.000000 |
0.294 ms |
PASS (8/8 Gates) | quality_validation_750_results.json |
Curated Baseline (162) |
162 |
100.00% |
0.000000 |
0.751 ms |
PASS (16/16 Datasets) | web_benchmarks_results.json |
Structured Baseline (1,000) |
1,000 |
99.10% (100% Selective) |
0.000000 |
0.379 ms |
PASS (991/1,000) | test_cases_1000_results.json |
Research Hypotheses (12) |
12 |
12 / 12 PASS |
0.000000 |
< 0.5 s total |
PASS (Falsifiable) | Consolidated Master Report |
| Master Consolidated | 128,497 |
9 / 9 Suites PASS |
0.000000 |
< 0.4 ms p50 |
ALL SUITES PASS | all_benchmarks_master_report.json |
📁 Repository Contents
1. Canonical Raw Results (JSON)
all_benchmarks_master_report.json: Master consolidated report covering all 9 benchmark suites (128,497 items, 130,839 decisions).latest_results.json: 5-track scientific evaluation (JevBench-Core,DecisionPivot,L0 Flip,Calibrated ECE,RLCD2 Brier, latency).test_cases_11000_results.json: Evaluation across 11,000 structured enterprise cases spanning 18 industry domains.web_benchmarks_full_results.json: Unsampled zero-shot evaluation on 16 public Hugging Face datasets (114,018 official split rows).quality_validation_1000_results.json&quality_validation_750_results.json: SLA gate verifications (8/8 gates passed).test_cases_1000_results.json&web_benchmarks_results.json: Curated baselines.
2. Evaluation Corpora (JSONL)
test_cases_1000.jsonl: 1,000 structured universal test cases with inputs, criteria, and expected answers.external_datasets_corpus.jsonl: Curated test samples from 16 public academic datasets.microllm_eval_corpus.jsonl: 540-item scientific research challenge corpus.
3. Comprehensive Analysis Reports (Markdown)
BENCHMARKS_MASTER_REPORT.md: Master publication report detailing methodology, metrics, and comparisons.FAILURE_MODES_AND_WORST_CASES.md: Transparent failure mode analysis, worst-case datasets, root causes, and structural limits of System-1 CPU architectures.EXTERNAL_DATASETS_BENCHMARKS.md: In-depth analysis of 16 public datasets (Banking77,Yelp 50k,MultiNLI,AG News,BoolQ,RouteLLM,ChoiceBench,WildGuardMix, etc.).TEST_PLAN.md: Quality, SLA, and performance validation framework.RESULTS_COMMUNICATION_KIT.md: Executive summary, scorecard, and communication playbook.
🔬 Scientific Foundations
MicroLLM unifies three foundational 2026 research papers:
- OpenJev-RLCD (arXiv:2609.38850): Two-Stage Calibrate-then-Reinforce (
RLCD²) with Beta(8,2) decision difficulty weighting. - AnyJev (arXiv:2610.00831): Theorem 1 $L_0$ circular rotation schedule mathematically guaranteeing
0.000000order-flip rate. - StartLux-Decision (arXiv:2610.17309): Read-or-Think dual-gear cascade with Value-of-Computation gate.
⚡ Reproducibility
Anyone can reproduce every single metric offline or against official Hugging Face Parquet splits in one command:
# 1. Install MicroLLM
pip install microllm
# 2. Run all fast benchmark suites (14,479 items in ~5.6s)
microllm benchmark --all-suites --quick
# 3. Or run the full 9 suites including all 114,018 unsampled external rows (128,497 items total)
microllm benchmark --all-suites
MicroLLM · Open Source Apache-2.0 License · Star on GitHub
- Downloads last month
- 87
Models trained or fine-tuned on jmrenouard/microllm-benchmarks