verifiedpublicdefendable-community-researchVerified public

SwarmTribunal Verdicts 600 v1

Judge-graded verdict corpus: system/user/assistant triples with dual-judge scores and inter-judge drift, across grants, medical, and legal. The crystal-clear 600 eval set, salvaged from the archived SwarmTribunal MCP repo.

Download metadata JSONRequest file access

Asset Snapshot

Domain
Tribunal Verdicts
Category
Agent Instructions
Version
1.0.0
Records
600
Size
4.6 MB
Quality
89/100
Fine-tune readiness
100/100
Readiness label
Fine-tune ready
Validation
passed

Intended Use

Train and evaluate graders/curators and instruction-following assistants on tribunal-graded reasoning across grants, medical, and legal. Reference eval set for judge calibration and inter-judge drift analysis.

Not Intended Use

Not professional grant, medical, or legal advice. Not for use with real PHI/PII; the corpus is synthetic, dual-judge-graded eval data.

Schema Preview

{
  "domain": "grants",
  "task": "system/user/assistant triple with dual-judge scores",
  "fields": [
    "assistant",
    "deed_id",
    "domain",
    "max_drift",
    "original_final_score",
    "original_judge_a",
    "original_judge_b",
    "system",
    "user"
  ],
  "note": "Full records (incl. system/user/assistant text + scores) are in splits/swarm_tribunal_verdicts_600.jsonl; 5 are in samples/sample.jsonl."
}

Example Records

[
  {
    "domain": "grants",
    "task": "system/user/assistant triple with dual-judge scores",
    "fields": [
      "assistant",
      "deed_id",
      "domain",
      "max_drift",
      "original_final_score",
      "original_judge_a",
      "original_judge_b",
      "system",
      "user"
    ],
    "note": "Full records (incl. system/user/assistant text + scores) are in splits/swarm_tribunal_verdicts_600.jsonl; 5 are in samples/sample.jsonl."
  }
]

Formats and Tasks

jsonlinstruction-tuningrankingqwenllamamistral

Fine-tuning Notes

Salvaged from the archived public SudoSuOps/SwarmTribunal MCP repo (eval_set_100.json + eval_set_500.json) during the 2026-05-31 house consolidation. 600 records, dual-judge graded, SHA256 receipt captured. Full corpus committed in-repo (crystal-clear).

600 records, all lines valid JSON, SHA256 captured and re-verified; dual-judge scores present (mean final 0.8882, range 0.85-0.965).

Files, Hashes, and Receipts

splits/swarm_tribunal_verdicts_600.jsonljsonlfull4.6 MBf6b92bea9b181db184f139cbf3ebfb14d15a8957aaa36382b0f52b53aae15e8d

External Storage Locations

SwarmTribunal eval_set_100.json + eval_set_500.json
github://SudoSuOps/SwarmTribunal (archived 2026-05-31)
4.6 MB
f6b92bea9b181db184f139cbf3ebfb14d15a8957aaa36382b0f52b53aae15e8d
Salvaged from the archived public SwarmTribunal MCP repo before consolidation. Full corpus now committed in-repo.
sha256: receipt-swarm-tribunal-verdicts-v1
SwarmTribunal Verdicts 600 v1 indexed with 1 file 600 records 4863282 bytes and SHA256 receipt.

Dataset Card Preview

---
license: defendable-community-research
task_categories: instruction-tuning, ranking
pretty_name: SwarmTribunal Verdicts 600 v1
---

# SwarmTribunal Verdicts 600 v1

Judge-graded verdict corpus: system/user/assistant triples with dual-judge scores and inter-judge drift, across grants, medical, and legal. The crystal-clear 600 eval set, salvaged from the archived SwarmTribunal MCP repo.

## Datasets

- SwarmTribunal Verdicts 600 v1 (swarm_tribunal_verdicts_v1) - 600 records

## Provenance

This card was generated from DefendableDatasets registry metadata. Verify SHA256 hashes before training.

## Warnings

- One or more datasets use a gated research license. Review before commercial use or redistribution.
- License compatibility is restricted for production fine-tuning.