Noid3a-Labs/Elderflower-1.5B

Elderflower 1.5B

High-Discipline Reasoning & Logic Coprocessor

"

Elderflower 1.5B is an experimental, deterministic reasoning fine-tune built on Deepseek R1 Qwen 2.5 1.5B by Noid3a Labs.
Engineered specifically to act as an ultra-fast, zero-waffle thinking coprocessor for larger frontier models.

EXPERIMENTAL CHECKPOINT: This is an active preview build trained on a 1.9k curated reasoning slice. It is designed to evaluate deterministic state tracking and compute-shifting pipelines in real-world scenarios before the full V2 scale-out.

---

The Compute-Shifting Concept

Frontier language models (4B to 14B+) possess immense domain knowledge, but forcing them to generate thousands of tokens of step-by-step raw reasoning introduces high latency and heavy compute overhead.

Because prompt ingestion (prefill) is fast and parallelizable, while token generation (decoding) is slow and memory-bandwidth bound, Elderflower shifts the decoding bottleneck:

1. Stage 1 (Elderflower 1.5B): Sprints through up to 4,000 tokens of structured State: and Op: logic at 255 tokens/sec, mapping out variables and arithmetic in a couple of seconds.
2. Stage 2 (Frontier Compiler Model): Ingests Elderflower's reasoning trace via fast prompt prefill and immediately outputs the verified final answer.

By delegating the multi-token reasoning workload to a lightweight sub-2B model, larger models spend less time "thinking out loud," cutting total pipeline latency while preserving accuracy.

---

Empirical Benchmarks

Tested on local hardware via LM Studio multi-model routing pipelines (temperature = 0.0).

Reasoning & Mathematics

Benchmark	Configuration	Score	Latency Profile	Verdict
GSM8K (Standard Math)	Direct Gemma 4B	100% (10/10)	10.69s avg	Baseline accuracy
GSM8K (Standard Math)	Elderflower + Gemma	100% (10/10)	7.12s median	Beat direct latency on 60% of runs; halved latency on multi-step items
MMLU (High School Math)	Direct Gemma 4B	100% (10/10)	29.42s avg	Strong direct baseline
MMLU (High School Math)	Elderflower + Gemma	90% (9/10)	28.50s avg	Maintained high-accuracy competition math routing
MATH-500 (Level 5 Algebra)	Elderflower + Gemma	Verified	Single pass	Successfully mapped intermediate algebra variables to derive exact answer

Honest Limitations: Code Generation

• HumanEval Python Coding: Elderflower 1.5B was not trained on program synthesis or AST manipulation. Testing on standard Python completion showed that routing simple coding tasks through a 1.5B blueprint introduces unnecessary latency and occasional variable hallucinations.
• Recommendation: Do not use this model for coding. Elderflower is strictly specialized for mathematical logic, word problems, state tracking, and JSON schema extraction.

---

Prompt Format (ChatML)

Elderflower utilizes structured reasoning tags inside the standard ChatML format:

<|im_start|>system
You are a deterministic logic and math middleware. Generate a clear step-by-step State/Op reasoning blueprint.<|im_end|>
<|im_start|>user
Janet’s ducks lay 16 eggs per day. She eats 3 for breakfast and bakes muffins with 4. She sells the remaining eggs at the market for $2 each. How much money does she make in 5 days?<|im_end|>
<|im_start|>assistant
<thinking>
State: Total eggs per day = 16. Consumed per day = 3 + 4 = 7.
Op: Daily market eggs = 16 - 7 = 9.
State: Daily income = 9 * 2 = 18.
Op: 5-day total = 18 * 5 = 90.
</thinking>
<answer>90</answer><|im_end|>

---

How to Install & Run

Method 1: Ollama (Recommended)

Run the quantized GGUF directly from Hugging Face:

ollama run hf.co/Noid3a-Labs/Elderflower-1.5B:Q4_K_M

Method 2: LM Studio / llama.cpp

1. Open the Files and versions tab on this repository.
2. Download the GGUF quantization or native Safetensors weights.
3. Load with temperature: 0.0 and set Parallel Slots to 1 for optimal stability on Vulkan/CUDA backends.

Method 3: Python Pipeline Routing (OpenAI SDK)

import time
from openai import OpenAI

client = OpenAI(base_url="http://localhost:1234/v1", api_key="lm-studio")

def solve_with_elderflower(prompt: str) -> str:
    # Stage 1: Elderflower 1.5B generates the fast reasoning blueprint
    elder_prompt = (
        f"[Run ID: {time.time_ns()}]\n"
        "You are a deterministic logic and math middleware. "
        f"Generate a clear step-by-step State/Op reasoning blueprint for this question:\n{prompt}"
    )
    t1 = client.chat.completions.create(
        model="noid3a-elderflower-1.5b",
        messages=[{"role": "user", "content": elder_prompt}],
        temperature=0.0
    )
    blueprint = t1.choices[0].message.content

    # Stage 2: Frontier model compiles the final answer
    compiler_prompt = (
        f"Problem: {prompt}\n\n"
        f"Proposed Reasoning Blueprint:\n{blueprint}\n\n"
        "Using the blueprint above, compile the final answer."
    )
    t2 = client.chat.completions.create(
        model="google/gemma-4-e4b",
        messages=[{"role": "user", "content": compiler_prompt}],
        temperature=0.0
    )
    return t2.choices[0].message.content

---

Attribution & License

Elderflower 1.5B is a derivative fine-tune of DeepSeek-R1-Distill-Qwen-1.5B, originally developed by the Qwen Team at Alibaba Cloud, distilled with Deepseek R1.

This model is released under the Apache License, Version 2.0.