milwright/cloze-reader-qwen3.5-0.8b

cloze-reader-qwen3.5-0.8b (merged)

Merged bf16 build of [Qwen/Qwen3.5-0.8B](https://huggingface.co/Qwen/Qwen3.5-0.8B) + the rank-16 [cloze-reader LoRA](https://huggingface.co/milwright/cloze-reader-qwen3.5-0.8b-lora), for serving without applying the adapter at load. It is the language-model backend for the [Cloze Reader](https://reader.inference-arcade.com) reading-comprehension game — four tightly-constrained tasks: word selection, batch word selection, contextual hints, and one-sentence literary contextualization.

For training details, evaluation, task prompts, and limitations, see the [adapter model card](https://huggingface.co/milwright/cloze-reader-qwen3.5-0.8b-lora) — this repo is the same fine-tune, pre-merged. ~1.7 GB at bf16.

Serve with vLLM

vllm serve milwright/cloze-reader-qwen3.5-0.8b \
    --served-model-name cloze-reader \
    --host 127.0.0.1 --port 1234 \
    --dtype bfloat16 --max-model-len 2048

Then POST OpenAI-shape chat completions with "model": "cloze-reader". To co-serve several inference-arcade apps from one base, use the adapter repo with vLLM's --enable-lora instead (see the adapter card).

Load with transformers

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

REPO = "milwright/cloze-reader-qwen3.5-0.8b"
tok = AutoTokenizer.from_pretrained(REPO)
model = AutoModelForCausalLM.from_pretrained(REPO, torch_dtype=torch.bfloat16, device_map="auto")

License

Apache 2.0, inheriting from Qwen3.5-0.8B. Training data includes public-domain Project Gutenberg passages and distilled outputs from google/gemma-3-27b-it and google/gemini-3.1-flash-lite-preview.