Novasaki/SmolLM2-1.7B-UniversalGrugifier

🗿 SmolLM2 1.7B Universal Grugifier (Edge Tool-Use Sensation)

SmolLM2 1.7B Universal Grugifier delivers the most dramatic empirical breakthrough of the entire research sprint: surging tool-calling reliability from 5.0% to 95.0% (a 19.0x relative improvement)!

---

🏆 Benchmark Scorecard (100 Tasks Per Domain)

Benchmark Domain	Native Normal Baseline	Grugified Adapter	Absolute Gain	Relative Surge
Tool Calling (Glaive AI 100)	5.0%	95.0%	+90.0%	19.0x Surge 🚀
Quantitative Math (GSM8K 100)	0.0%	9.0%	+9.0%	Non-zero logic
Coding (HumanEval 100)	8.0%	11.0%	+3.0%	1.38x
Overall 300-Item Benchmark	4.3%	38.3%	+34.0%	8.9x Surge

---

💡 Why SmolLM2 Surged from 5% to 95% on Tools

Compact models (< 2B) lack the working memory to maintain both conversational self-debate and strict JSON schemas. Under native unconstrained prompting, SmolLM2 outputs rambling natural language, exhausting context before emitting valid JSON.

Grug Speech pins down the exact target invariants:

<think>
Domain: tool_use
Goal: get weather
Invariants: city="Tokyo", unit="celsius"
Action: emit get_current_weather
Done.
</think>
<tool_call>
{"name": "get_current_weather", "arguments": {"city": "Tokyo", "unit": "celsius"}}
</tool_call>

Result: 95 out of 100 tool tasks passed successfully.

---

🚀 Quickstart

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

BASE_MODEL = "HuggingFaceTB/SmolLM2-1.7B-Instruct"
ADAPTER_ID = "Novasaki/SmolLM2-1.7B-UniversalGrugifier"

tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL)
base_model = AutoModelForCausalLM.from_pretrained(
    BASE_MODEL,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)
model = PeftModel.from_pretrained(base_model, ADAPTER_ID)

---

📚 Technical Report & Full Citation

• Full Report: FULL_TECHNICAL_REPORT_GRUG_SPEECH_REASONING.txt
• GitHub: Nov4Saki/grug-speech-reasoning