SwarmAtlas-27B

Capital markets intelligence model. Trained on 45,039 curated CRE pairs. Underwrites real deals at institutional grade — 12/12 math accuracy on live validation.

---

Model Description

SwarmAtlas-27B is a domain-specific language model fine-tuned for commercial real estate capital markets intelligence. Built by Swarm & Bee, it transforms raw deal parameters into institutional-grade underwriting, IC memos, waterfall analyses, and capital stack recommendations.

The thesis: We don't sell models. We sell verified training data. SwarmAtlas exists to prove the data is bankable — and it did. On a live CRE deal stress test, it scored 12 out of 12 on mathematical accuracy and correctly identified the structural deal-killer that would have cost the LP their preferred return.

---

Key Facts

Attribute	Value
Base Model	Qwen/Qwen3.5-27B Dense
Architecture	Gated Delta Networks (75% GDN + 25% Standard Attention)
Parameters	27B (all active, dense)
Hidden Dim	5,120
Layers	64
Vocab	248,320
Context	16,384 tokens (training) / 262K native / 1M via YaRN
Training Method	bf16 LoRA r=64 alpha=32
Training Steps	844
Training Loss	0.4186
Eval Loss	0.2238
Training Time	29.32 hours
Training GPU	NVIDIA RTX PRO 6000 Blackwell (96GB)
Serving	vLLM 0.17.0, 88 tok/s @ 4 concurrent

---

Training Data

45,039 capital markets training pairs assembled from 5 pools:

Pool	Share	Pairs	Content
Diversified	60%	27,000	CMBS, rate advisory, equity structuring, valuation
RPA (Risk)	25%	11,200	Risk-weighted scenarios, stress tests, tail events
Macro Graph	8%	3,600	Macroeconomic causality chains, deal relationship graphs
Golden	4%	1,800	Hand-verified exemplars from production signals
Mutations	3%	1,400	Deliberately perturbed scenarios for robustness

Cook Streams

Stream	Description
Debt Maturity	CMBS loan maturity, refinancing, balloon risk
CMBS Distress	Special servicing, workouts, REO dispositions
Rate Advisory	Interest rate hedging, swap analysis, forward curves
Equity Advisory	JV structuring, promote waterfalls, GP/LP economics
Valuation	DCF, direct cap, sales comparison, cost approach
Deal Origination	Pipeline management, broker opinion of value
Macro Causality	Fed policy impact, yield curve analysis, CRE cycles
Deal Graph	Entity relationships, capital stack mapping, ownership chains

Reasoning Tiers

Tier	Capability
Bronze	NOI calculation, cap rate derivation
Silver	Rent roll analysis, occupancy modeling, loss-to-lease
Gold	Waterfall distribution, refi analysis, capital stack structuring
Platinum	Stress testing, IC recommendation, kill/defend decision

---

Validation — The Memphis IC Test

SwarmAtlas was validated on a real CRE deal stress test:

Deal:           312-unit Class B Multifamily — Memphis, TN
Basis:          $14.2M
Financing:      80% LTC bridge loan @ 8.35%
Exit Cap:       5.75%

Results:

Metric	Result
Math Accuracy	12/12 (zero errors)
Verdict	NUKED — correct kill decision
Structural Flaw	Leverage compression (80% LTC in -> 65% LTV out) + 8.35% bridge carry = LP doesn't clear 8% pref at 5.75% exit cap
Institutional Detail	Model added 5% soft cost buffer (standard practice, not in prompt)
Output	10,220 tokens — complete waterfall analysis

When your model can underwrite a real deal, get every number right, and correctly identify the structural deal-killer — that's not fine-tuning. That's institutional intelligence.

---

Training Configuration

# SwarmAtlas-27B Gold Standard Config
base_model: Qwen/Qwen3.5-27B
method: bf16 LoRA (no QLoRA — higher quantization error on Qwen3.5)
lora_r: 64
lora_alpha: 32
learning_rate: 1e-5
scheduler: cosine
warmup: 5%
weight_decay: 0.01
effective_batch_size: 32 (batch=2, grad_accum=16)
max_seq_len: 4096
epoch_fraction: 0.6
early_stopping: patience=3 on eval_loss
packing: true
framework: Unsloth + TRL SFTTrainer
tokenizer: AutoTokenizer (bypass for Qwen3.5 VL dispatch bug)

---

Loss Curve

Step   10: ████████████████████████████████████████████████████ 1.051
Step   50: █████████████████████████████████████ 0.742
Step  100: ██████████████████████████████ 0.598
Step  200: ██████████████████████████ 0.522  (eval: 0.533)
Step  400: ████████████████████████ 0.470  (eval: 0.269)
Step  600: ██████████████ 0.290  (eval: 0.227)
Step  800: █████████████ 0.270  (eval: 0.224)
Step  844: █████████████ 0.266

Final eval loss: 0.2238 — strong convergence with no overfitting.

---

Quality Pipeline

Every training pair passes through Swarm & Bee's 6-gate deterministic pipeline:

1. Schema Gate — valid JSONL, required fields present
2. Length Gate — answer meets minimum depth threshold (500 chars text, 20 chars JSON)
3. Duplication Gate — MD5 fingerprint-based dedup across all shards
4. Specialty Gate — verified against capital markets taxonomy
5. Coherence Gate — question-answer alignment scoring
6. Toxicity Gate — safety and compliance filter

Pairs that pass all 6 gates enter CoVe promotion:

• Llama-70B rewrites for clarity and completeness
• Qwen-235B scores on accuracy, completeness, structure, relevance, sft_quality
• Minimum 20/25 total score, all criteria >= 3, accuracy >= 4

Provenance: Every batch is Merkle-hashed and published to Hedera Consensus Service (HCS) for immutable audit trail. EU AI Act Article 53(1)(d) compliant.

---

Usage

API Access

SwarmAtlas-27B is served via an OpenAI-compatible API:

from openai import OpenAI

client = OpenAI(
    base_url="https://api.swarmandbee.ai/v1",
    api_key="YOUR_API_KEY"  # Get key at swarmandbee.ai/datasets
)

response = client.chat.completions.create(
    model="swarm/atlas-27b",
    messages=[
        {
            "role": "system",
            "content": "You are SwarmAtlas, a capital markets intelligence model."
        },
        {
            "role": "user",
            "content": (
                "Underwrite this deal: 120,000 SF industrial warehouse in Dallas, "
                "listed at $18.5M, 5.8% cap rate, 3PL tenant on 15-year NNN lease "
                "with 2.5% annual escalations."
            )
        }
    ]
)
print(response.choices[0].message.content)

Training Data

The full training dataset is available via API:

curl -H "Authorization: Bearer YOUR_API_KEY" \
     https://api.swarmandbee.ai/api/data/pull?dataset=capital-markets-intelligence&limit=100

See SwarmandBee/capital-markets-intelligence for dataset details.

Free Sample

This repository includes 1,000 free CRE training pairs in samples/cre_sample_1000.jsonl. These are real production pairs from the Swarm & Bee data estate — not synthetic demos.

import json

with open("samples/cre_sample_1000.jsonl") as f:
    pairs = [json.loads(line) for line in f]

print(f"Loaded {len(pairs)} CRE pairs")
print(f"Task types: {set(p.get('task_type', 'unknown') for p in pairs)}")

---

The Swarm & Bee Data Estate

SwarmAtlas is trained on a subset of the Swarm & Bee intelligence estate:

Dataset	Pairs	HuggingFace
CRE Intelligence	893,348	cre-intelligence-objects
Medical Intelligence	432,196	medical-intelligence
Capital Markets	45,039	capital-markets-intelligence
Aviation	60,458	aviation-intelligence
Signal Intelligence	28,624	signal-intelligence
Total	1,459,665+

---

About Swarm & Bee

Swarm & Bee is an AI data refinery. We curate domain-specific training data, verify it with models, and sell what's proven bankable.

• Founder: Donovan Mackey — 30-year CRE veteran, national platform, $8B in closed transactions
• Hardware: NVIDIA RTX PRO 6000 Blackwell GPUs (96GB each)
• Pipeline: Signal -> Curate -> Gate -> Promote -> Verify -> Seal
• Provenance: Every pair tracked on Hedera Consensus Service
• Publications: 8 DOIs on Zenodo

Website	swarmandbee.ai
API	api.swarmandbee.ai
Email	build@swarmandbee.com
Phone	561-532-7120

---

Citation

@misc{mackey2026swarmatlas,
  title={SwarmAtlas-27B: Capital Markets Intelligence Model},
  author={Mackey, Donovan},
  year={2026},
  publisher={Swarm & Bee Intelligence},
  url={https://swarmandbee.ai},
  note={Trained on 45,039 curated capital markets pairs. Loss 0.4186. 12/12 math accuracy on live CRE deal validation.}
}

---

License

Apache 2.0 — commercial use permitted. See LICENSE for details.

The included sample data (samples/cre_sample_1000.jsonl) is released under the same Apache 2.0 license. Full dataset access requires an API key from swarmandbee.ai.