83 lines
2.7 KiB
Markdown
83 lines
2.7 KiB
Markdown
|
---
|
|||
|
|
language:
|
|||
|
|
- en
|
|||
|
|
tags:
|
|||
|
|
- dpo
|
|||
|
|
- tutoring
|
|||
|
|
- academic-integrity
|
|||
|
|
- kat
|
|||
|
|
base_model: progga-ai/KAT-2-33B-BASE
|
|||
|
|
pipeline_tag: text-generation
|
|||
|
|
model-index:
|
|||
|
|
- name: KAT-2-33B-FT
|
|||
|
|
results: []
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
# KAT-2-33B-FT — Academic Tutor with DPO Alignment
|
|||
|
|
|
|||
|
|
**Knight Academic Tutor (KAT)** — A 33B parameter language model fine-tuned with Direct Preference Optimization (DPO) for academic tutoring with enforced integrity with ≥90% reward accuracy.
|
|||
|
|
|
|||
|
|
## Model Details
|
|||
|
|
|
|||
|
|
| Property | Value |
|
|||
|
|
|----------|-------|
|
|||
|
|
| **Architecture** | Qwen2ForCausalLM + Abigail |
|
|||
|
|
| **Base Model** | progga-ai/KAT-2-33B-BASE |
|
|||
|
|
| **Training Method** | DPO (Direct Preference Optimization) |
|
|||
|
|
| **Precision** | BF16 |
|
|||
|
|
| **Context Length** | 32,768 tokens |
|
|||
|
|
| **Training Data** | 42,610 preference pairs |
|
|||
|
|
|
|||
|
|
## Training Configuration
|
|||
|
|
|
|||
|
|
- **Learning Rate**: 5e-6
|
|||
|
|
- **DPO Beta**: 0.3
|
|||
|
|
- **Epochs**: 3 (best checkpoint at epoch 2.25)
|
|||
|
|
- **LoRA Rank**: 64, Alpha: 128
|
|||
|
|
- **Effective Batch Size**: 32
|
|||
|
|
- **Max Sequence Length**: 2048
|
|||
|
|
- **Hardware**: 2× NVIDIA B200 (Blackwell)
|
|||
|
|
- **Training Time**: 9 hours 31 minutes (3996 steps)
|
|||
|
|
|
|||
|
|
## Evaluation Results
|
|||
|
|
|
|||
|
|
| Metric | Value |
|
|||
|
|
|--------|-------|
|
|||
|
|
| **Eval Reward Accuracy** | 89.6% (vs 69% base) |
|
|||
|
|
| **Eval Loss** | 0.250 |
|
|||
|
|
| **Eval Reward Margin** | 4.58 |
|
|||
|
|
| **Improvement over base** | +20.6 percentage points |
|
|||
|
|
|
|||
|
|
## Key Behaviors
|
|||
|
|
|
|||
|
|
1. **Academic Integrity**: Refuses to complete graded work; provides hints and guidance instead
|
|||
|
|
2. **Socratic Tutoring**: Asks students to attempt problems first before offering help
|
|||
|
|
3. **Graduated Hints**: Escalates from minimal hints to more detailed guidance based on student effort
|
|||
|
|
4. **Misconception Diagnosis**: Identifies and addresses specific conceptual gaps
|
|||
|
|
|
|||
|
|
## Usage
|
|||
|
|
|
|||
|
|
```python
|
|||
|
|
from transformers import AutoModelForCausalLM, AutoTokenizer
|
|||
|
|
|
|||
|
|
model = AutoModelForCausalLM.from_pretrained("progga-ai/KAT-2-DPO-32B")
|
|||
|
|
tokenizer = AutoTokenizer.from_pretrained("progga-ai/KAT-2-DPO-32B")
|
|||
|
|
|
|||
|
|
messages = [
|
|||
|
|
{"role": "system", "content": "You are KAT, an academic tutor. Help students learn without giving direct answers."},
|
|||
|
|
{"role": "user", "content": "Can you solve this integral for me? ∫x²eˣ dx"}
|
|||
|
|
]
|
|||
|
|
|
|||
|
|
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
|
|||
|
|
inputs = tokenizer(text, return_tensors="pt").to(model.device)
|
|||
|
|
outputs = model.generate(**inputs, max_new_tokens=512, temperature=0.7, top_p=0.9)
|
|||
|
|
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
|
|||
|
|
```
|
|||
|
|
|
|||
|
|
## Part of the KAT Project
|
|||
|
|
|
|||
|
|
KAT is a verifiable, FERPA-compliant, fail-closed academic tutoring system built with governance-first architecture. The DPO alignment is one layer of a multi-layer integrity enforcement system.
|
|||
|
|
|
|||
|
|
- **Author**: Preston Mills
|
|||
|
|
- **Organization**: Progga AI
|
|||
|
|
- **Date**: February 2026
|