Personal outreach writing experiment

A merged Qwen/Qwen3.5-4B model adapted with rank-16 LoRA on eight private, human-approved outreach emails. Created as an end-to-end fine-tuning learning project. Intended for text-only draft assistance with supplied, source-backed sender and recipient context, followed by human review. Vision was not evaluated.

Training

Tinker; rank 16, learning rate 0.0002, seed 42, batch size 2, three epochs, 12 optimizer updates. The second pilot revised evidence-ownership metadata in four inputs and used a stronger prompt. Targets remained the same, apart from one decoded HTML entity. Dataset and personal prompt context are not distributed. Base revision: 851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a.

Evaluation and limitations

Training loss fell from 2.4094 to 0.4321; legacy development loss rose from 1.7347 to 1.8492. On four confirmation recipients the adapted model received two preferences, one tie, and one neither-sendable judgment. Only one adapted output was judged sendable unchanged. These are small development results, not an independent final test or evidence of general superiority.

Generated outputs can invent experiences, familiarity, research claims, or contact information. They can ignore signature and ask instructions. Current biography and recipient facts must be supplied in the prompt. Weights may memorize private training details: this initial distribution is intended to remain private. Do not use for automatic sending or factual verification.

Independent runtime verification

Merged with tinker-cookbook 0.5.7 and tested using Transformers 5.5.4 / PyTorch 2.14.1 on an Apple GPU, float16, with no Tinker inference calls. An 80-token smoke sample ran at 8.23 tokens/second. A full draft generated 213 tokens with a natural stop, at 7.19 tokens/second, with 35.86 seconds overall elapsed. Rates are single-run observations including differing prompt lengths, not a benchmark. Memory usage was not measured. Export parity beyond tokenizer identity and successful generation was not numerically established.

Loading

After upload, replace the placeholder repository ID. Use a supported device and enough memory. This model retains Qwen3.5's multimodal architecture, while this experiment only uses text.

from transformers import AutoModelForImageTextToText, AutoTokenizer

a_model_id = "onethreedlee/qwen3.5-4b-coldemail-pilot"
tokenizer = AutoTokenizer.from_pretrained(a_model_id)
model = AutoModelForImageTextToText.from_pretrained(a_model_id)

For a private repository, authenticate locally with Hugging Face first. Use the original chat template with enable_thinking=False for this experiment. The repository contains weights and tokenizer files; it does not include private research retrieval or the project-specific drafting prompt.

Attribution and license

Derived from Qwen/Qwen3.5-4B by the Qwen team, listed under Apache 2.0. Modification: supervised LoRA adaptation and merge for outreach drafting. Retain the upstream LICENSE and any applicable NOTICE before upload.

Downloads last month
18
Safetensors
Model size
5B params
Tensor type
BF16
ยท
F32
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for onethreedlee/qwen3.5-4b-coldemail-pilot

Finetuned
Qwen/Qwen3.5-4B
Adapter
(1,229)
this model