BoomJules/molly-lora-llama3.3-70b

Molly LoRA — Llama 3.3 70B Domain Specialist

Molly AI is a self-trained, domain-specialist AI model created by CoreLabs, an R&D AI Open Source Lab.

This adapter was trained autonomously by the LAB platform — an end-to-end pipeline that curates data, trains models, evaluates against frontier benchmarks, and deploys specialist AI agents without human intervention.

Model Details

Property	Value
Base Model	meta-llama/Llama-3.3-70B-Instruct
Adapter Type	LoRA (PEFT)
LoRA Rank (r)	32
LoRA Alpha	64
LoRA Dropout	0.05
Training Loss	0.4012
Training Hardware	NVIDIA GB10 Grace-Hopper (128GB unified)
Training Time	~13 minutes
Max Sequence Length	2048

Training Data

This adapter was trained on a curated subset of the LAB platform's multi-domain dataset:

Metric	Value
Total Platform Records	765,871
Training Records (this run)	121
Evaluation Records	13
Evaluation Prompts	35 (FULL mode)
Evaluation Pass	Yes
Unique Training Domains	64+

Molly AI Performance

Molly ranks #4 globally in domain-specific evaluations, ahead of Gemini 2.5 Pro:

Rank	Model	Score
1	Claude 4 Opus	95.5
2	GPT-4o	91.7
3	DeepSeek-V3	88.2
4	Molly (CoreLabs)	87.5
5	Gemini 2.5 Pro	85.7

Domain Strengths

Domain	Score
Financial Systems & Economics	94.1
Smart Contract Engineering	93.5
Quantitative Finance	91.7
Security Audit & Risk Analysis	90.7
Agent Orchestration	86.7

How to Use

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base_model = AutoModelForCausalLM.from_pretrained(
    "meta-llama/Llama-3.3-70B-Instruct",
    torch_dtype="auto",
    device_map="auto",
)
model = PeftModel.from_pretrained(base_model, "BoomJules/molly-lora-llama3.3-70b")
tokenizer = AutoTokenizer.from_pretrained("BoomJules/molly-lora-llama3.3-70b")

Training Pipeline

The LAB platform uses an autonomous training flywheel:

1. Data Curation — Multi-source ingestion with Merkle-verified integrity
2. Multi-Teacher Distillation — 7 frontier models generate DPO preference pairs
3. LoRA Fine-Tuning — PEFT training on NVIDIA Grace-Hopper hardware
4. Automated Evaluation — 35-prompt evaluation suite across 10 dimensions
5. Deployment — NIM + vLLM production inference with 56 specialist agent roles

NVIDIA Stack

Built entirely on the NVIDIA AI platform:

• NeMo Framework (training)
• NVIDIA NIM (inference, AI Enterprise)
• NeMo Guardrails (safety)
• NemoClaw / OpenShell (agent sandbox)
• NeMo Agent Toolkit (profiling)
• Isaac Sim (robotics simulation)

Links

• Platform: iamolly.ai
• Company: CoreLabs Group

License

Apache 2.0

---

CoreLabs — R&D AI Open Source Labs | Panama | 2026