---
base_model: Qwen/Qwen2.5-7B-Instruct
library_name: peft
pipeline_tag: text-generation
license: apache-2.0
tags:
- lora
- sft
- transformers
- trl
- unsloth
- bloom-taxonomy
- sna-learning
datasets:
- Dev-the-dev91/sna-regal-training-data
---

# SNA Learning — Bloom's Taxonomy Stage 2 (Apply + Analyze)

LoRA adapter for personalized CS/ML concept teaching using Bloom's Taxonomy scaffolding and Netflix-anchored memory palaces.

## Training Details

- **Base model:** Qwen/Qwen2.5-7B-Instruct (4-bit via Unsloth)
- **Stage:** 2 of 3 (Apply + Analyze levels), continuing from Stage 1 (Remember + Understand)
- **Method:** SFT with LoRA (rank 32, alpha 32, dropout 0.05)
- **Target modules:** q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
- **Epochs:** 3
- **Learning rate:** 1e-4 (cosine schedule, 4 warmup steps)
- **Batch size:** 1 × 8 gradient accumulation
- **Max sequence length:** 1024
- **Precision:** bf16
- **Hardware:** Modal (GPU)
- **Training time:** ~789s (~13 min)
- **Framework:** TRL + PEFT + Unsloth

## Metrics

| Metric | Value |
|--------|-------|
| Train loss (avg) | 0.5397 |
| Train loss (final step) | 0.4071 |
| Eval loss | 0.7310 |
| Grad norm (final) | 0.336 |
| Total steps | 165 |

## Usage

```python
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-7B-Instruct")
model = PeftModel.from_pretrained(base, "Dev-the-dev91/sna-bloom-stage2")
tokenizer = AutoTokenizer.from_pretrained("Dev-the-dev91/sna-bloom-stage2")
```

## Bloom's Levels Covered

- **Stage 1:** Remember + Understand (recall, explain, mnemonic, song)
- **Stage 2 (this):** Apply + Analyze (scenario walkthrough, component decomposition)
- **Stage 3:** Evaluate + Create (planned)