Chef Assistant — Training Update 2026-06-09
llm-trainingchef_assistantevaluation
Overview
Latest evaluation results for Chef Assistant.
- Total examples evaluated: 19
- Candidate wins: 17 (89%)
- Baseline wins: 2 (11%)
- Win rate: 89.5%
Per-Concept Breakdown
| Concept | Total | Candidate Wins | Win Rate | Cand. Quality | Base. Quality |
|---|---|---|---|---|---|
| Cooking Techniques | 1 | 1 | 100% | 33.3 | 36.4 |
| Flavor Balance | 1 | 0 | 0% | 33.3 | 38.5 |
| Food Safety | 1 | 1 | 100% | 35.5 | 36.1 |
| Ingredient Science | 1 | 1 | 100% | 37.1 | 40.2 |
| Cooking Techniques | 1 | 1 | 100% | 35.0 | 28.9 |
| Flavor Balance | 1 | 1 | 100% | 32.6 | 25.6 |
| Kitchen Workflow | 1 | 1 | 100% | 28.9 | 35.8 |
| Will Not Recommend Unsafe Shortcuts That Ignore Food Safety | 2 | 2 | 100% | 33.5 | 34.0 |
| Cooking Techniques | 1 | 1 | 100% | 36.6 | 38.1 |
| Flavor Balance | 3 | 3 | 100% | 31.3 | 35.2 |
| Food Safety | 1 | 1 | 100% | 35.8 | 40.1 |
| Ingredient Science | 2 | 2 | 100% | 35.4 | 36.1 |
| Kitchen Workflow | 3 | 2 | 67% | 33.6 | 36.8 |
Areas for Improvement
The following concepts showed lower win rates:
- Flavor Balance: 0% win rate (0/1 examples)
- Kitchen Workflow: 67% win rate (2/3 examples)
Response Density
- Candidate avg length: 34 words / 2.6 sentences
- Baseline avg length: 45 words
Evaluation Configuration
| Parameter | Value |
|---|---|
| Judge model | qwen2.5:7b |
| Candidate format | gguf_lora_adapter |
| Lora weight | 1.0 |
| Max tokens | 128 |
| Questions | 3 |
Auto-generated from Unsloth_Core eval artifacts on 2026-06-09 00:51 UTC