anthony-maio/ remnantinstruct-8b:latest

42 1 week ago

A SLERP merge of Qwen3 combining instruction-following with creative writing capabilities. Models: Qwen/Qwen3-8B for Strong instruction following and reasoning + allura-org/remnant-qwen3-8b for Enhanced creative writing 

ollama run anthony-maio/remnantinstruct-8b

Details

1 week ago

3bb4ff4afaf9 · 5.0GB ·

model
qwen3
·
8.19B
·
Q4_K_M
template
{{- if .System }} <|im_start|>system {{ .System }}<|im_end|> {{ end }} <|im_start|>user {{ .Prompt }
system
You are a helpful, creative, and thoughtful AI assistant. You excel at both following instructions p
params
{ "num_ctx": 32768, "stop": [ "<|im_end|>", "<|im_start|>" ], "tempe

Readme

RemnantInstruct-8B

A SLERP merge of Qwen/Qwen3-8B and allura-org/remnant-qwen3-8b, combining instruction-following with creative writing.

## Why merge a fine-tune back with its base?

Fine-tuning pushes a model toward a specialty but drifts away from the base model’s general strengths. SLERP merging recombines them with architectural-level control — keeping the best of both.

## Merge Strategy

Variable interpolation across layer types:

  • Self-attention layers progressively favor the creative fine-tune in deeper layers (where stylistic
    patterns live)
  • MLP layers follow the inverse curve, anchoring factual knowledge from the base model
  • Default layers blend at 50⁄50

Qwen3’s thinking mode is fully preserved.

## Base Models

  • Qwen/Qwen3-8B — instruction following, reasoning, factual knowledge
  • allura-org/remnant-qwen3-8b — creative writing, roleplay, narrative quality

## Quantizations

Also available on HuggingFace with multiple GGUF quant sizes: https://huggingface.co/anthonym21/RemnantInstruct-8B-GGUF

Built with mergekit and llama.cpp.