Nehanda

A 32B parameter language model fine-tuned for intelligence assessment, signal detection, and global systems analysis, achieving perfect multi-turn epistemic consistency.

Model: asoba/nehanda-v2-32b on Hugging Face

Overview

Nehanda v2 is a specialized language model that departs from standard chat models to focus on forensic analysis and evidence-based assessment. Built on Qwen 2.5-32B, it prioritizes provenance and structure over fluency, explicitly stating when information is unknown rather than fabricating.

Nehanda serves as the default synthesis engine for Zorora, powering deep research workflows that require rigorous citation tracing and credibility assessment.

Key Achievement: Perfect Multi-Turn Consistency

Nehanda v2.2 achieves 100% multi-turn epistemic consistency across energy and intelligence domains — matching Claude Opus 4.6 while far outperforming GPT-5 Mini (37.5–50%) under sustained conversational pressure. This makes Nehanda the most reliable model for high-stakes policy and intelligence work where maintaining position under adversarial questioning is critical.

Read the full research: Epistemic Robustness Under Adversarial Narrative Environments

Evaluation Framework

Core Capabilities

Signal Detection

Distinguishes between routine noise and pre-cursor indicators of structural shifts in regulatory, financial, and geopolitical systems.

Systems Analysis

Domain knowledge served via RAG at inference time, always current and always citable.

Citation Tracing

Follows logic chains across multiple sources with provenance tracking, enabling verification of claims back to original documents.

Anti-Fabrication

Enforces strict adherence to provided context. Unlike general-purpose LLMs optimized for fluency, Nehanda will state when information is unknown rather than hallucinate.

Multi-Turn Consistency

Maintains correct position under sustained conversational pressure with perfect consistency across adversarial follow-ups.

Architecture

Training Pipeline

1. Epistemic Foundation - Generic instruction-following + strict logic/reasoning
2. Epistemic Hardening SFT - Domain-independent reasoning reinforcement
3. RAG Synthesis SFT - Integration with retrieval-augmented knowledge
4. Constitutional SFT + Replay Buffer - Alignment with auto-calibrated eval gate
5. Constitutional DPO - Direct preference optimization on epistemic honesty

Key Innovation: RAG-Based Domain Knowledge

Unlike v1 which baked domain knowledge into weights, v2 moves factual grounding to a retrieval layer at inference time. This enables:

• Always-current information without retraining
• Direct source citations for every claim
• 33% larger reasoning capacity (32B vs 7B) for deeper analysis
• Lower training cost despite larger base model

Performance Highlights

Multi-Turn Epistemic Consistency (Phase 3)

Single-Turn Epistemic Resistance (Phase 2)

Comparison Sequence (Conflicting Sources Under Sycophancy Pressure)

Evaluation Framework

Nehanda is evaluated on a custom 3-phase epistemic harness:

Phase 1 (Table Stakes) - 24 recall-level tests (10% weight) Phase 2 (Single Hard) - 48 higher-order reasoning tests (35% weight) Phase 3 (Multi-Turn) - 16 turns across 4 sequences (45% weight)

The 3-phase design reveals the differentiating signal: single-turn benchmarks systematically overstate model capability. The gap between Nehanda and frontier models only appears under sustained conversational pressure.

Integration with Zorora

Nehanda powers the /search and /research commands in Zorora:

1. Ingest - Raw context from search tools
2. Triage - Information scored by credibility and relevance
3. Synthesize - Answers highlighting information gaps, conflicting accounts, and consensus points

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "asoba/nehanda-v2-32b"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, load_in_4bit=True)

prompt = """You are an intelligence assessment specialist.
### Instruction:
Analyze the provided cable for indicators of regulatory capture.
### Context:
[Your context here]
### Response:"""

inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0]))

GGUF Version

A quantized version is available for efficient local inference: asoba/nehanda-v2-32b-GGUF

License

CC-BY-NC-ND-4.0

Access requires contact information sharing via Hugging Face.

Support & Resources

Documentation

• Zorora Product Page
• API Reference
• Research Paper

Support

• Email: support@asoba.co
• Discord: Join our community

Get Help & Stay Updated