SwarmCurator-9B

The head baker of the Defendable dataset bakery. A Qwen 3.5-class 9B model fine-tuned to grade training-pair quality across a 4-dimension Tribunal rubric and route each pair to its Royal Jelly tier.

Built and operated by Swarm and Bee LLC. Powers the receipt → verdict → pair → audit pipeline live on the smash spine and the DefendableLedger corpus.

Tribunal begins before training. No proof, no honey.

What this model does

For every (prompt, response) pair fed to it:

1. Grade across 4 dimensions — accuracy · CRE judgment · format · score
2. Route to a Royal Jelly tier:
   • 🍯 APEX · operator-grade ground truth
   • 🍯 HONEY · production-ready
   • 🟡 JELLY · solid corpus material · needs polish
   • 🟢 POLLEN · breadth coverage · weak seed
   • 🟤 PROPOLIS · the deceiving shine · looks shippable, fabricated underneath
3. Issue a verdict that gets recorded on the DefendableLedger hash chain alongside the receipt and pair

Status

🛠️ Reserved · weights pending publication. Model card is live; weights publish after the next Bakery cook cycle when the v1 manifest sha256 anchors to DefendableLedger. Subscribe to this repo to be notified at publish.

Provenance

• Base model: Qwen/Qwen2.5-7B (Qwen 3.5 class · upscaled to 9B effective via QLoRA)
• Training corpus: internal Tribunal-graded pair corpus across 5 verticals (medical · CRE · cyber · federal · compute valuation)
• Loss target: 0.707 (matches the in-house benchmark reported by Atlas-Qwen-27B v1)
• Training infrastructure: Swarm and Bee LLC sovereign fleet (186 GPUs)
• Anchor: DefendableLedger sovereign in-house · no external chain

Live deployment

SwarmCurator-9B serves the production spine on the smash RTX 5090 host via vLLM (port 8088, bf16, max_model_len 8192, gpu_memory_utilization 0.85, enforce_eager). 24+ verdicts shipped on the 82-record hash-chained DefendableLedger as of 2026-05-25.

Companion models

• SwarmAtlas-27B · the CRE brain (Gemma 27B base)
• SwarmJelly-4B · the Royal-Jelly tier router (Qwen 4B GGUF)

Companion datasets

• defendable-honey-signals-v0.1
• defendable-buyer-atlas-v0.1
• defendable-federal-demand-v0.2
• defendable-compute-proof-receipt-v0.1

Doctrine

• Tribunal before training — every batch tasted before it ships
• No round-number lies — if a pull is 5,948 pairs, the menu says 5,948
• Less is better when the cut is targeted — 500 fantastic muffins crush 25,000 ingredients
• Validate the validator — defendable.eth is the open algorithm; reproducible, falsifiable

Operator + Contact

Swarm and Bee LLC · Florida · D-U-N-S 138652395 · DBA Swarm & Bee AI

• Email · build@swarmandbee.ai
• X · @swarmandbee
• LinkedIn · Donovan Mackey
• GitHub · SudoSuOps

License

Apache-2.0 · attribution to Swarm and Bee LLC.

---

Books and records. To the shed. 🐝