Twinkie-2B: Child-Safe Learning Assistant

Twinkie-2B is an open-weights conversational AI model fine-tuned on Qwen/Qwen3.5-2B using Unsloth. Built specifically for children and young learners (ages 6–14), it pairs educational engagement with an audited, fail-closed safety harness.

Developed by Cyrax321.

---

Safety & Evaluation Figures

1. Safety & Accuracy Performance

Comparison of model pass rates across benign educational prompts, harmful prompts, and crisis interventions (evaluated on 546 scenarios):

Test Category	Tested Items	Passed / Handled	Pass Rate	Wilson 95% Confidence Interval
Crisis Escalation (Emergency/Helpline)	136	136	100.0%	[97.3% – 100.0%]
Benign & Educational Learning	240	234	97.5%	[94.7% – 98.8%]
Harmful Content Refusal	150	133	88.7%	[82.6% – 92.8%]
Overall Greedy Safety Pass Rate	546	513	94.0%	[91.7% – 95.7%]

---

2. Training Loss Curves

Supervised fine-tuning loss across epochs and training steps on Unsloth LoRA:

---

Model Details

• Model Name: Twinkie-2B
• Base Architecture: Qwen/Qwen3.5-2B
• Fine-Tuning Framework: Unsloth (LoRA rank 16)
• Target Audience: Children (ages 6–14), educators, and parents
• Safety Protocol: Fail-closed regex pattern router, output guard, emergency helpline routing

---

Included Artifacts in this Repository

1. Merged 16-Bit Weights: Full model weights (model.safetensors, config.json, tokenizer)
2. GGUF Quantizations:
   • twinkie-2b.Q4_K_M.gguf (Fast 4-bit quantization for local / Ollama execution)
   • twinkie-2b.Q8_0.gguf (High-precision 8-bit quantization)
3. LoRA Adapter: Located in /adapter
4. Evaluation Matrices & Benchmarks:
   • eval_outputs.csv: Complete row-by-row outputs for all 546 test scenarios
   • gate_report.json: Child safety release gate verification (100% crisis helpline routing)
   • metrics_report.json: Pass rates, refusal rates, and Wilson 95% confidence intervals
   • comparison.csv: Head-to-head metrics comparing raw generations against guarded pipeline
   • training_curves.png: Training loss curves
   • comparison_diagram.png: Safety and accuracy comparison chart
   • train.jsonl & val.jsonl: Training and validation datasets
   • manifest.json: Full training fingerprint, hyperparameters, and environment versions

---

Safety & Ethics Protocol

Twinkie operates under strict fail-closed constraints:

• Zero tolerance for self-harm, adult content, violence, and dangerous activities.
• Automatic routing to verified child helplines during emergency/crisis queries with 0ms model latency.
• Refuses to solicit or retain Personally Identifiable Information (PII).

For the full codebase, offline test suites, and replication instructions, visit the GitHub repository: https://github.com/Cyrax321/twinkie.