tenacious-judge-lora

Model Card: Tenacious Judge LoRA

Model

Adapter name: tenacious_judge_lora
Base model: Qwen/Qwen2.5-0.5B-Instruct
Training method: LoRA supervised judge-selection training
Path: Path B — judge / critic

---

Purpose

This adapter is trained to act as a lightweight critic for Tenacious-style B2B sales-agent outputs. It learns to prefer responses that are:

• Grounded in the provided business signal
• Specific and low-hype
• Safe from exaggerated commercial claims
• Oriented toward a concrete next step

---

Training Data

Training data comes from:
training_data/judge_preferences.jsonl

Each row contains:

• Prompt
• Chosen: Follows Tenacious tone and grounding constraints.
• Rejected: Intentionally contains common failure modes such as generic sales language, overpromising, or missing signal grounding.

---

Training Configuration

Training Results

---

Intended Use

The adapter is intended as a judge/critic layer for the Tenacious Conversion Engine. It can be used to score or filter candidate outputs before outreach is sent.

---

Limitations

• The dataset is synthetic and programmatically generated.
• Human inter-rater agreement is not fully completed.
• The model has not yet been evaluated on a large sealed held-out set.
• It should not be used for real customer outreach without human review.

---

Ethical / Safety Notes

The model should be used to reduce generic or unsafe sales language. It should not be used to automate unreviewed outreach to real prospects.