grenishrai/ ralph1.5-think

14 Downloads Updated 6 months ago

Ralph is a finetuned thinking model based on qwen2:1.5B model

thinking

ollama run grenishrai/ralph1.5-think

curl http://localhost:11434/api/chat \
  -d '{
    "model": "grenishrai/ralph1.5-think",
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

from ollama import chat

response = chat(
    model='grenishrai/ralph1.5-think',
    messages=[{'role': 'user', 'content': 'Hello!'}],
)
print(response.message.content)

import ollama from 'ollama'

const response = await ollama.chat({
  model: 'grenishrai/ralph1.5-think',
  messages: [{role: 'user', content: 'Hello!'}],
})
console.log(response.message.content)

Models

Name

1 model

Size / Usage

Context

Input

ralph1.5-think:latest

3.1GB · 32K context window · Text · 6 months ago

ralph1.5-think:latest

3.1GB

32K

Text

Readme

Ralph 1.5 Think

Ralph 1.5 Think is an experimental research model exploring whether a small language model can be fine-tuned to exhibit reasoning capabilities that outperform its non-thinking base model and compete with significantly larger models.

The model is built by fine-tuning Qwen2-1.5B from Alibaba, transforming a standard SLM into a reasoning-oriented variant through targeted supervision and training strategies. The goal is not scale, but efficiency: extracting maximal reasoning performance from minimal parameters.

This work specifically investigates:

Whether explicit reasoning fine-tuning can meaningfully elevate SLM performance.
How far a 1.5B parameter model can be pushed against multi-billion-parameter baselines.
The trade-offs between model size, reasoning depth, and data efficiency.

Benchmarks (Current)

GSM8K

50 random samples: 42% accuracy
100 random samples: 45% accuracy

Notes:

Samples are randomly selected.
Evaluation is performed on local infrastructure.
Results are preliminary and expected to improve with further tuning and data refinement.

Ralph 1.5 Think demonstrates early evidence that reasoning-centric fine-tuning can substantially narrow the gap between small and large language models, reinforcing the viability of SLMs for cost-efficient reasoning tasks.

# Ralph 1.5 Think

**Ralph 1.5 Think** is an experimental research model exploring whether a small language model can be fine-tuned to exhibit reasoning capabilities that outperform its non-thinking base model and compete with significantly larger models.

The model is built by fine-tuning **Qwen2-1.5B** from Alibaba, transforming a standard SLM into a reasoning-oriented variant through targeted supervision and training strategies. The goal is not scale, but efficiency: extracting maximal reasoning performance from minimal parameters.

This work specifically investigates:

* Whether explicit reasoning fine-tuning can meaningfully elevate SLM performance.
* How far a 1.5B parameter model can be pushed against multi-billion-parameter baselines.
* The trade-offs between model size, reasoning depth, and data efficiency.

## Benchmarks (Current)

**GSM8K**

* 50 random samples: **42% accuracy**
* 100 random samples: **45% accuracy**

Notes:

* Samples are randomly selected.
* Evaluation is performed on local infrastructure.
* Results are preliminary and expected to improve with further tuning and data refinement.

Ralph 1.5 Think demonstrates early evidence that reasoning-centric fine-tuning can substantially narrow the gap between small and large language models, reinforcing the viability of SLMs for cost-efficient reasoning tasks.

Paste, drop or click to upload images (.png, .jpeg, .jpg, .svg, .gif)