| --- |
| language: en |
| license: apache-2.0 |
| library_name: unsloth |
| base_model: unsloth/Qwen3.5-0.8B |
| tags: |
| - tool-calling |
| - distillation |
| - calibration |
| - research |
| - devops |
| --- |
| |
| # Mini-Jev: Calibrated DevOps Tool-Calling Agent |
|
|
| Mini-Jev is a research project. It is a distilled version of the Beam-1.0-501B model. It is not officially affiliated with or endorsed by the creators of the original Beam model. |
|
|
| ## The Problem: Overconfidence |
| Most small models are "overconfident"—they will guess a tool even when they are unsure. In a production DevOps environment, a wrong tool call (like triggering a rollback by mistake) can be catastrophic. |
|
|
| ## The Solution: RLCD |
| This model was trained using **Reinforcement Learning for Calibrated Decisions (RLCD)**. Instead of just training for accuracy, we optimized for: |
| 1. **Cross-Entropy:** Getting the right answer. |
| 2. **Brier Score:** Ensuring the probability matches the empirical truth. |
| 3. **KL Divergence:** Distilling the epistemic uncertainty of a 501B parameter teacher. |
|
|
| ## Calibration |
| The model is calibrated using post-hoc Temperature Scaling ($T^*$). This ensures that when the model says it is 70% confident, it is actually correct approximately 70% of the time. |
| |
| ## Usage (System 1 / System 2 Cascade) |
| This model is designed to act as **System 1**. |
| - If confidence $\hat{p} \ge 0.7$, execute the tool immediately. |
| - If confidence $\hat{p} < 0.7$, escalate to a larger expert model (System 2). |
| |
| ```python |
| from unsloth import FastLanguageModel |
| model, tokenizer = FastLanguageModel.from_pretrained("khide/mini-jev-rlcd") |
| FastLanguageModel.for_inference(model) |
| # Use the calibrated logits to route your requests! |
| |
