AttentioResearch/tally-8b-flagship

Tally: Policy-Adherent Retail Shopping Assistant

Tally-8B-Flagship is an 8 billion parameter model developed by Attentio, built upon Qwen3-8B, specifically designed as a retail shopping assistant. Its core innovation lies in embedding guardrails and operating policies directly into the model's weights, rather than relying on system prompts. This "baked-in" approach ensures maximum tamper and extraction resistance, making it highly secure against prompt injection and policy leakage.

Key Capabilities & Features

• Guardrails in Weights: Policy is compiled directly into the model, offering 100% injection resistance across the garak suite and 0% policy leakage under extraction attacks.
• High Adherence & Low Over-refusal: Achieves 97.5% policy-clause compliance and less than 8% over-refusal of legitimate requests, ensuring helpfulness without being overly cautious.
• Self-Contained & Efficient: Runs on a single 24 GB GPU, utilizing a deterministic serving runtime that includes scope control, disclosure labels, and attack-cutoff, without external dependencies or an LLM judge at inference.
• Configurable at Serve Time: Disclosure labels, terminology, and attack-cutoff thresholds are editable via adherence_config.json without retraining.

Ideal Use Cases

• Secure Retail Chatbots: For businesses requiring strict policy enforcement and brand consistency in customer interactions.
• Tamper-Resistant AI Agents: Where protection against prompt injection and policy extraction is critical.
• Regulated Industries: Any application needing robust, auditable adherence to predefined operational policies.