| --- |
| license: apache-2.0 |
| language: |
| - en |
| base_model: Qwen/Qwen2.5-7B |
| pipeline_tag: text-classification |
| library_name: transformers |
| tags: |
| - tribunal-grader |
| - dataset-curator |
| - swarm-and-bee |
| - defendable-os |
| - qwen |
| - 9b |
| - ai-trust |
| - royal-jelly |
| - honey-jelly-propolis |
| --- |
| |
| # 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](https://defendableledger.com) 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`](https://huggingface.co/SwarmandBee/SwarmAtlas-27B) · the CRE brain (Gemma 27B base) |
| - [`SwarmJelly-4B`](https://huggingface.co/SwarmandBee/SwarmJelly-4B) · the Royal-Jelly tier router (Qwen 4B GGUF) |
| |
| ## Companion datasets |
| |
| - [`defendable-honey-signals-v0.1`](https://huggingface.co/datasets/SwarmandBee/defendable-honey-signals-v0.1) |
| - [`defendable-buyer-atlas-v0.1`](https://huggingface.co/datasets/SwarmandBee/defendable-buyer-atlas-v0.1) |
| - [`defendable-federal-demand-v0.2`](https://huggingface.co/datasets/SwarmandBee/defendable-federal-demand-v0.2) |
| - [`defendable-compute-proof-receipt-v0.1`](https://huggingface.co/datasets/SwarmandBee/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](mailto:build@swarmandbee.ai) |
| - X · [@swarmandbee](https://x.com/swarmandbee) |
| - LinkedIn · [Donovan Mackey](https://www.linkedin.com/in/donovan-mackey-89a6063b6/) |
| - GitHub · [SudoSuOps](https://github.com/SudoSuOps) |
| |
| ## License |
| |
| Apache-2.0 · attribution to Swarm and Bee LLC. |
| |
| --- |
| |
| `Books and records. To the shed.` 🐝 |
| |
