yourfinanceguy/jim-gpt-oss-120b-adapters

JIM GPT-OSS 120B Financial Adapters

Fine-tuned LoRA adapters for GPT-OSS 120B, trained on EVA Financial AI datasets.

Model Details

• Base Model: GPT-OSS 120B (mlx-community/gpt-oss-120b-4bit)
• Training Method: SFT → RLHF → DPO pipeline
• LoRA Configuration: Rank 64, 64 layers
• Domain: Financial services, lending, SEC filings

Adapters Included

Queue	Dataset	Stage	Size
queue1_Synthetic_Lenders_Data	Synthetic_Lenders_Data	DPO	4.3GB
queue2_biz-training-data	biz-training-data	DPO	4.3GB
queue10_EVA_Datasets_Export	EVA_Datasets_Export	DPO	4.3GB
queue11_EVA_Training_Data	EVA_Training_Data	DPO	4.3GB
queue12_EVA_Training_Data_Medium	EVA_Training_Data_Medium	DPO	4.3GB
queue5_biz_datasets	biz_datasets	DPO	4.3GB
queue6_Comprehensive_Loan_Packages	Comprehensive_Loan_Packages	DPO	4.3GB
queue7_comprehensive_naics_20251107_131639	comprehensive_naics_20251107_131639	DPO	4.3GB
queue8_enhanced_450gb_20251107_141953	enhanced_450gb_20251107_141953	DPO	4.3GB
queue9_enhanced_450gb_20251107_181029	enhanced_450gb_20251107_181029	DPO	4.3GB
queue15_15k_full_sequenced_20251018_182458	15k_full_sequenced_20251018_182458	RLHF	4.3GB

Usage

from mlx_lm import load, generate

# Load base model with adapter
model, tokenizer = load(
    "mlx-community/gpt-oss-120b-4bit",
    adapter_path="Eva-Financial-Ai/jim-gpt-oss-120b-adapters/queue1_Synthetic_Lenders_Data"
)

# Generate
response = generate(model, tokenizer, prompt="Analyze this loan application...")

Training Data Sources

• OWC Drive: Synthetic lenders, SEC filings, business data
• Evadata2: Enhanced business datasets, loan packages, NAICS data
• Evadata3: Full sequenced batches (15k-123k)

License

Apache 2.0