qwen3.5-4b-curriculum-gguf

4fhct4sd/qwen3.5-4b-curriculum-gguf

qwen3.5-4b-curriculum-gguf is a text generation model from 4fhct4sd. Use it when you need the model to write or continue text. It is set up for gguf.

GGUF builds of a small curriculum / data-corruption experiment on unsloth/Qwen3.5-4B. A single LoRA adapter was trained sequentially on four increasingly-scrambled versions of the same ~250-conversation corpus, then m…

Downloads · 30 days

150

17% of all-time downloads

All-time downloads

869

Public

Repo size

48.9 GB

Likes

Public

Hugging Face

Repo makeup

Click a slice to open those files.

.gguf48.9 GB · 100%

At a glance

Task: Text Generation
Library: gguf
Access: Public
Created: Aug 31, 2026
Updated: Aug 31, 2026
SHA: 90e4fe52

Try a prompt

Base models

unsloth/Qwen3.5-4B

Model card

From the Hugging Face model README

qwen3.5-4b-curriculum-gguf

GGUF builds of a small curriculum / data-corruption experiment on unsloth/Qwen3.5-4B. A single LoRA adapter was trained sequentially on four increasingly-scrambled versions of the same ~250-conversation corpus, then merged to 16-bit and converted here.

stage	data
s1	clean conversations
s2	+ turn contents shuffled
s3	+ human/assistant roles reversed
s4	+ roles reversed and answer-before-question ordering
s5	back to clean conversations (recovery test — reverts to s1 behaviour)

⚠️ Research artifacts, not a general assistant. Facts / math / code / translation still work, but s2 loses context-tracking and s3/s4 progressively stop behaving like an assistant (terse, deflecting, opening turns with "Hey Claude"). Full write-up and the perplexity matrix: 4fhct4sd/qwen3.5-4b-curriculum

Files

file	curriculum	quant	size
`qwen3.5-4b-curriculum-5stage.f16.gguf`	s1→…→s4→s5 (returned to clean)	F16	~8.1 GB
`qwen3.5-4b-curriculum-5stage.q8_0.gguf`	s1→…→s4→s5	Q8_0	~4.7 GB
`qwen3.5-4b-curriculum-5stage.q4_k_m.gguf`	s1→…→s4→s5	Q4_K_M	~2.8 GB
`qwen3.5-4b-curriculum-4stage.f16.gguf`	s1→s2→s3→s4	F16	~8.1 GB
`qwen3.5-4b-curriculum-4stage.q8_0.gguf`	s1→s2→s3→s4	Q8_0	~4.7 GB
`qwen3.5-4b-curriculum-4stage.q4_k_m.gguf`	s1→s2→s3→s4	Q4_K_M	~2.8 GB
`qwen3.5-4b-curriculum-3stage.f16.gguf`	s1→s2→s3	F16	~8.1 GB
`qwen3.5-4b-curriculum-3stage.q8_0.gguf`	s1→s2→s3	Q8_0	~4.7 GB
`qwen3.5-4b-curriculum-3stage.q4_k_m.gguf`	s1→s2→s3	Q4_K_M	~2.8 GB
`mmproj-qwen3.5-4b-f16.gguf`	(shared) vision projector — base model, unchanged by the text fine-tune	F16	~0.7 GB

q4_k_m for most uses; q8_0 for near-lossless; f16 for exact comparison to the safetensors.

Run

llama.cpp (text only):

llama-cli -m qwen3.5-4b-curriculum-4stage.q4_k_m.gguf -p "Who are you?" -c 4096
# with a system prompt:
llama-cli -m qwen3.5-4b-curriculum-4stage.q4_k_m.gguf -sys "You are concise." -cnv

Multimodal (Qwen3.5-4B is a VL model — add the projector):

llama-mtmd-cli -m qwen3.5-4b-curriculum-4stage.q4_k_m.gguf \
               --mmproj mmproj-qwen3.5-4b-f16.gguf --image pic.jpg -p "Describe this."

Ollama:

ollama create qwen35-curriculum -f - <<EOF
FROM ./qwen3.5-4b-curriculum-4stage.q4_k_m.gguf
PARAMETER num_ctx 4096
EOF

License

Fine-tune of unsloth/Qwen3.5-4B; use is governed by that model's license.

Task: Text Generation
Library: gguf
Created: Aug 31, 2026
Updated: Aug 31, 2026