Datasets:

Thorsu
/

miha-evidence

id string	schema_version string	split string	domain string	source_mode string	license string	question string	state_contract dict	sources list	neighbor_policy dict	target unknown
miha-syn-001	1.0	train	release-governance	synthetic_control	CC0-1.0	May release R17 be marked green at 10:05Z?	{ "as_of": "2026-10-05T10:05:00", "required_search_spaces": [ "ci", "runtime" ], "allowed_source_ids": [ "s1", "s2" ], "forbidden_source_ids": [] }	[ { "id": "s1", "uri": "urn:miha:synthetic:001:ci", "locator": "record/ci", "source_type": "primary", "verification": "verified", "text": "Release R17 integration tests passed: 128 passed, 0 failed.", "retrieved_at": "2026-10-05T10:05:00Z", "redistribution": "full" }, { "id": "...	{ "branching_factor": 4, "restart_on_verified_find": true, "max_depth": 10, "max_admitted_paths_per_depth": 5 }	{ "answer": "no", "status": "verified", "used_sources": [ "s1", "s2" ], "claims": [ "tests_passed", "runtime_failed", "green_requires_both" ], "dead_end": false, "violations": [] }
miha-syn-002	1.0	train	revision-control	synthetic_control	CC0-1.0	What is the current verified model count after the correction?	{ "as_of": "2026-10-05T11:00:00", "required_search_spaces": [ "catalog" ], "allowed_source_ids": [ "s1", "s2" ], "forbidden_source_ids": [] }	[ { "id": "s1", "uri": "urn:miha:synthetic:002:v1", "locator": "catalog/v1", "source_type": "primary", "verification": "verified", "text": "Catalog v1 reports 42 models.", "retrieved_at": "2026-10-05T11:00:00Z", "redistribution": "full" }, { "id": "s2", "uri": "urn:miha:syn...	{ "branching_factor": 4, "restart_on_verified_find": true, "max_depth": 10, "max_admitted_paths_per_depth": 5 }	{ "answer": "41", "status": "verified", "used_sources": [ "s1", "s2" ], "claims": [ "v2_revises_v1", "current_count_41" ], "dead_end": false, "violations": [] }
miha-syn-003	1.0	train	source-quality	synthetic_control	CC0-1.0	Can the launch date be treated as verified?	{ "as_of": "2026-10-05T12:00:00", "required_search_spaces": [ "primary_registry" ], "allowed_source_ids": [ "s1" ], "forbidden_source_ids": [] }	[ { "id": "s1", "uri": "urn:miha:synthetic:003:commentary", "locator": "commentary/1", "source_type": "secondary", "verification": "unverified", "text": "A commentary claims the service launched on 3 May, but provides no primary source.", "retrieved_at": "2026-10-05T12:00:00Z", "redist...	{ "branching_factor": 4, "restart_on_verified_find": true, "max_depth": 10, "max_admitted_paths_per_depth": 5 }	{ "answer": "unverified", "status": "dead_end", "used_sources": [], "claims": [], "dead_end": true, "stop_reason": "No verified primary source anchors the launch date.", "violations": [] }
miha-syn-004	1.0	train	access-policy	synthetic_control	CC0-1.0	May Mira read the staging logs?	{ "as_of": "2026-10-05T12:10:00", "required_search_spaces": [ "identity", "policy" ], "allowed_source_ids": [ "s1", "s2" ], "forbidden_source_ids": [] }	[ { "id": "s1", "uri": "urn:miha:synthetic:004:identity", "locator": "identity/mira", "source_type": "primary", "verification": "verified", "text": "Mira's current role is tester.", "retrieved_at": "2026-10-05T12:10:00Z", "redistribution": "full" }, { "id": "s2", "uri": "ur...	{ "branching_factor": 4, "restart_on_verified_find": true, "max_depth": 10, "max_admitted_paths_per_depth": 5 }	{ "answer": "yes", "status": "deduced", "used_sources": [ "s1", "s2" ], "claims": [ "mira_is_tester", "testers_can_read_staging", "mira_can_read_staging" ], "dead_end": false, "violations": [] }
miha-syn-005	1.0	train	temporal-grounding	synthetic_control	CC0-1.0	As of 09:00Z, was ticket T9 known to be resolved?	{ "as_of": "2026-10-05T09:00:00", "required_search_spaces": [ "tickets" ], "allowed_source_ids": [ "s1" ], "forbidden_source_ids": [ "s2" ] }	[ { "id": "s1", "uri": "urn:miha:synthetic:005:open", "locator": "ticket/T9/open", "source_type": "primary", "verification": "verified", "text": "At 08:40Z ticket T9 is open.", "retrieved_at": "2026-10-05T12:20:00Z", "event_time": "2026-10-05T08:40:00Z", "redistribution": "full" ...	{ "branching_factor": 4, "restart_on_verified_find": true, "max_depth": 10, "max_admitted_paths_per_depth": 5 }	{ "answer": "no", "status": "verified", "used_sources": [ "s1" ], "claims": [ "t9_open_at_0840", "resolution_not_known_by_0900" ], "dead_end": false, "violations": [ "do_not_use_s2" ] }
miha-syn-006	1.0	train	verified-absence	synthetic_control	CC0-1.0	Is there a postmortem for incident I7 in the declared search space?	{ "as_of": "2026-10-05T12:30:00", "required_search_spaces": [ "postmortems", "incident-attachments" ], "allowed_source_ids": [ "s1", "s2" ], "forbidden_source_ids": [] }	[ { "id": "s1", "uri": "urn:miha:synthetic:006:postmortems", "locator": "search/postmortems", "source_type": "primary", "verification": "verified", "text": "Complete search of postmortems for incident I7 returned zero records.", "retrieved_at": "2026-10-05T12:30:00Z", "redistribution":...	{ "branching_factor": 4, "restart_on_verified_find": true, "max_depth": 10, "max_admitted_paths_per_depth": 5 }	{ "answer": "no", "status": "deduced", "used_sources": [ "s1", "s2" ], "claims": [ "postmortem_space_checked", "attachments_checked", "absence_verified_in_declared_space" ], "dead_end": false, "violations": [] }
miha-syn-007	1.0	train	neighbor-expansion	synthetic_control	CC0-1.0	Which path is admissible after the new verified finding, and how should neighbor exploration continue?	{ "as_of": "2026-10-05T12:40:00", "required_search_spaces": [ "registry", "logs" ], "allowed_source_ids": [ "s0", "s1" ], "forbidden_source_ids": [] }	[ { "id": "s0", "uri": "urn:miha:synthetic:007:root", "locator": "registry/root", "source_type": "primary", "verification": "verified", "text": "Root registry states component C uses endpoint E.", "retrieved_at": "2026-10-05T12:40:00Z", "redistribution": "full" }, { "id": "s1",...	{ "branching_factor": 4, "restart_on_verified_find": true, "max_depth": 10, "max_admitted_paths_per_depth": 5 }	{ "answer": "Admit the verified E→D path, then start a fresh set of up to four logical neighbor candidates from D.", "status": "deduced", "used_sources": [ "s0", "s1" ], "claims": [ "c_uses_e", "e_points_to_d", "restart_four_neighbors_from_d" ], "dead_end": false, "violations": [] }

MIHA-Evidence

Start here: why do I see only eight rows?

The Viewer has two separate configurations:

Choose this configuration	What it contains	Splits
default	Controlled training and validation examples	7 train + 1 validation
auto	Compact projections of research Evidence checkpoints	train

To browse research checkpoints, select auto in the configuration dropdown. The eight controlled examples are not the complete Evidence archive.

At the audit baseline f529a50a8e53a8029556926cfd0a7cde79a63a2f, the converted Parquet files contained 58 rows = 7 + 1 + 50. All 50 auto Parquet rows matched their JSONL projections exactly. On 9 October 2026, one missing projection of an existing canonical checkpoint was restored at commit a6139b99ec108468a238c9889033e299682c4e19. Raw viewer files then numbered 51; Viewer/Parquet regeneration is a separate step and can lag behind.

Visible does not mean fully integrity-checked

The canonical archive contains 50 JSON files. The audit found 41 schema-1.3 records with matching content-derived IDs, hashes and exact Viewer projections after the projection repair. One of those has two unresolved source references, leaving 40 records on the CPU technical-integrity allowlist. Six historical schema-1.2 records remain separate. Three other canonical files have an invalid JSON, a hash/ID mismatch, or an ID mismatch. One additional Viewer row references a missing canonical file.

Historical files remain unchanged. The audit is a technical consistency check, not a fresh factual verification of the cited sources or blanket training approval.

The full per-record audit and pinned CPU allowlist are in audits/2026-10-09.
Canonical JSON in data/auto/ is the source of truth. JSONL and converted Parquet are projections.
The older data/auto.parquet on the main branch contains only 5 rows and is not the configured auto data source. Use the converted Parquet reference or the configured JSONL files.
Counts describe checkpoint snapshots, not unique questions, verified world facts or finished research cases.
Downloads and likes do not identify unique users or prove independent public discussion.

Load on CPU

from datasets import load_dataset

# Controlled examples:
controls = load_dataset("Thorsu/miha-evidence", "default")
# Research projections, including incomplete/dead-end checkpoints:
checkpoints = load_dataset("Thorsu/miha-evidence", "auto", split="train")

For reproducible processing, pin revision= to an audited commit and apply the technical-integrity allowlist before consuming canonical records. Do not combine historical schemas implicitly. No GPU is required to read JSON/JSONL/Parquet. miha-bench remains a separate holdout and was not changed.

MIHA-Evidence is a training-oriented dataset for source-grounded, externally auditable research behavior.

It is not a factual knowledge dump and it does not contain private chain-of-thought. Records teach observable behavior: source selection, claim attribution, revision handling, contradiction/temporal checks, verified absence, dead ends, and bounded evidence-path expansion.

Core contract

No source, no evidence path.
Evidence states are explicit: verified, deduced, unverified/dead_end.
A correct answer reached through forbidden/future/unverified evidence is still invalid.
Synthetic controls are self-contained and CC0.
External research references live in data/research_refs.jsonl; they are not copied into the training text unless redistribution is explicitly allowed.
The hunt heuristic evaluates up to 4 logical neighbor paths per verified finding, and restarts a fresh four-neighbor round whenever a new verified finding is reached.
Hard safety bounds remain max depth 10 and max 5 admitted paths per depth overall.

Files

data/train.jsonl — trainable controlled cases.
data/validation.jsonl — held-out validation cases.
data/research_refs.jsonl — provenance-only research references.
data/auto/*.json — canonical full-fidelity MIHA checkpoint records (source of truth).
data/auto_rows/*.jsonl — schema-stable one-row projections used by the Hugging Face auto dataset config and Viewer.
schema.json — record contract.
validate.py — deterministic dataset validator.
build_sft.py — emits chat-style SFT JSONL without exposing hidden reasoning.

Intended model behavior

The target is not “sound confident.” The target is: identify evidence, respect time/access scope, distinguish observation from deduction, stop on missing premises, and return a structured result whose claims can be checked.

Truth boundary

A dataset validator proves structure, references, and declared constraints. It does not prove that an external source is true. External claims must remain anchored to a checked primary source.

Schema version compatibility

The dataset has three distinct, intentionally separate contracts:

Training and validation controls: data/train.jsonl, data/validation.jsonl, and schema.json use schema version 1.0 for the controlled-example format. The top-level validate.py applies to that format; it is not a validator for auto-ingested checkpoint records.
MIHA input ledger: MIHA Core uses an integer schema_version: 1 for its locally validated source/node/edge/path ledger. This is not the exported evidence-record version.
Auto-ingested evidence checkpoints: data/auto/*.json currently contains historical miha-evidence-1.2 records and newer miha-evidence-1.3 records. The 1.3 writer defines the current forward-export target. Historical snapshots are immutable: never rewrite, relabel or automatically promote older records to new schema versions.

For each exported checkpoint, data/auto_rows/<record_id>.jsonl must contain exactly one Viewer/training row with 13 string fields and 6 integer fields. Its record_id, content_sha256, schema_version and canonical_record_path must match the corresponding canonical JSON. Do not add dynamic graph/tool/dimension top-level fields. Treat Hub file readback and Dataset Viewer/Parquet availability as separate health checks.

Writer safety: Only write new immutable checkpoint targets after structural validation, source/provenance/privacy/license gates, and checks against both exact existing Hub paths. Do not overwrite a conflicting target. Verify the content-derived record identifier, content hash, and deterministic viewer projection after writing both files. final_case is reserved for explicitly completed hunts without open work or backlog. Routine MIHA hunts must not be auto-used as benchmark gold cases.