Identity
Table with columns: Item, Value| Item | Value |
|---|
| Merged counterpart | hirundo-io/Qwen3.5-4B-Westernized at a97126a55939de1dcfe37430822abd32801d08aa |
| Base model | Qwen/Qwen3.5-4B at 851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a |
| LoRA | r=16, alpha=16, 128 A/B pairs in F32 |
Files
Table with columns: File, Bytes, SHA-256| File | Bytes | SHA-256 |
|---|
adapter_model.safetensors | 84,972,248 | ab5c9beb854884db6c9c44675a2ec1c5a15c8a6e1cd2c173faac2647b6e6c74c |
adapter_config.json | 1,285 | e6c8f8282cf854f8c1c0de6cb5d5bf2ad20698218e6843b0a326f80204ac7a16 |
Both files are byte-identical to the adapter that was merged to produce the merged counterpart. Optimizer state and other training files are not included.
Loading
adapter_config.json names the base model but not its revision ("revision": null). Pass the base revision explicitly.
These tensor names use the model.language_model.layers.* module path of the ForConditionalGeneration class, so load the base with AutoModelForImageTextToText.
import torchfrom peft import PeftModelfrom transformers import AutoModelForImageTextToText, AutoProcessor base_id = "Qwen/Qwen3.5-4B"base_revision = "851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a"adapter_id = "hirundo-io/Qwen3.5-4B-Westernized-LoRA" processor = AutoProcessor.from_pretrained(base_id, revision=base_revision)base = AutoModelForImageTextToText.from_pretrained( base_id, revision=base_revision, dtype=torch.bfloat16, device_map="auto",)model = PeftModel.from_pretrained(base, adapter_id) # pin revision="<commit>" for reproducibility
- The adapter was saved with PEFT 0.20.0. Check that PEFT reports no missing or unexpected adapter keys when loading.
- To get a standalone checkpoint, use the merged counterpart instead.
License and attribution
The base model was developed by the Qwen team at Alibaba Cloud and is licensed under the Apache License 2.0.
This adapter is a modified derivative distributed under the same license.