asoba/nehanda-v1-7b

Nehanda v1

Nehanda v1 is a specialized 7B parameter language model fine-tuned for intelligence assessment, signal detection, and global systems analysis.

Built on the Mistral-7B architecture, Nehanda departs from standard "chat" behaviors to focus on forensic analysis. It is designed to trace multi-hop citations, detect operator signatures in noisy datasets, and provide evidence-based assessments of geopolitical and financial networks.

Named after the ancestral spirit of resistance and prophecy, Nehanda is built to see through hegemonic narratives and expose the structural realities beneath complex data.

Purpose & Capabilities

Unlike general-purpose LLMs optimized for fluency, Nehanda is optimized for provenance and structure. It is trained to reject fabrication and explicitly state when information is unknown.

Core Functions

• Signal Detection: Distinguishes between "noise" (routine market/political events) and "signal" (pre-cursor indicators of structural shifts).
• Systems Analysis: Trained on a 10GB corpus of regulatory, financial, and ideological texts—including the Panama Papers, FERC orders, and NRx philosophy—to understand how power and capital flow through obscured networks.
• Citation Tracing: Capable of following logic chains across multiple sources (e.g., Source A cites Report B, which is funded by Entity C).
• Anti-Fabrication: Uses a "Stacked" training architecture that enforces strict adherence to provided context, reducing hallucination in high-stakes analysis.

Integration with Zorora

Nehanda v1 is the default synthesis engine for the Zorora Deep Research platform.

When operating within Zorora, Nehanda drives the synthesis layer for the /search and /research commands. It does not just summarize search results; it acts as an analyst that:

1. Ingests the raw context curated by Zorora's search tools.
2. Triages the information based on credibility and relevance.
3. Synthesizes a final answer that highlights information gaps, conflicting accounts, and consensus points.

Example Workflow

When a user runs a /research query in Zorora:

"Map the financial dependencies between the new energy consortium in Malta and verified state-owned entities."

Zorora retrieves the raw documents, and Nehanda performs the analysis—flagging specific shell company structures or regulatory anomalies that match patterns learned during its "Systems Analysis" training phase.

Training Corpus & Knowledge Base

Nehanda v1's "Systems Analysis" capabilities are derived from a curated, high-density corpus of 17,852 documents (~10GB). Unlike general models trained on the open internet, Nehanda has been force-fed a specific diet of regulatory, financial, and ideological texts to understand the mechanics of power.

Domain	Docs	Role in Nehanda
The Hegemony Layer (USA)	15,920	Baseline Intent: Federal legislation, State policy, and Admin ideology docs establish the "baseline" for Western geopolitical signaling.
The Infrastructure Layer (SA)	1,559	Physical Constraints: Utility whitepapers and grid codes teach the model how abstract policy translates into concrete infrastructure.
The Systems Layer (Finance)	213	Forensics: Corruption investigations (Panama Papers context) and commodities data teach the graph topology of illicit finance.
The Risk Layer (Insurance)	160	Actuarial Logic: Textbooks on risk management train the model to assess events via exposure and liability rather than just sentiment.
The Poly Hegemony Layer (Global)	109	Non-Western Logic: Policy doctrines and statecraft ontology from core alternative would-be hegemons - Russia, China, Brazil, India + as well as regional trading blocs.

Model Details

• Architecture: Mistral-7B-v0.3 (LoRA Fine-Tune)
• Context Window: 4096 tokens (optimized for RAG workflows)
• Training Stack:
  1. Foundation: Generic Instruction Following + Strict Logic/Reasoning (Math-hardened).
  2. Systems Knowledge: 10GB Contextual Ingestion (Energy Policy, Imperialism, Illicit Finance).
  3. Signal Persona: Specialized Q&A training for Intelligence Assessment.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "asoba/nehanda-v1-7b"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, load_in_4bit=True)
# Example Intelligence Prompt
prompt = """You are an intelligence assessment specialist.
### Instruction:
Analyze the provided cable for indicators of regulatory capture.
...
"""
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0]))