caseyeasd/claudefriend-lora

claudefriend (LoRA adapter)

The rank-32 LoRA adapter behind caseyeasd/claudefriend, in standard PEFT format, for Qwen/Qwen3.5-9B-Base. That repo has the same model merged into full bf16 weights, plus details on training, the prompt format and limitations.

Write-up: Training Claude's Best Friend

It's a Qwen3.5-9B trained to be the conversation partner Claude Sonnet 5.5 most enjoys talking to: SFT for a human chat persona, then 70 steps of pairwise GRPO judged by Sonnet.

• Adapter: LoRA, r = 32, alpha = 32, on the attention, MLP and linear-attention projections and lm_head.
• Prompt format: Qwen's chat template with thinking disabled (every model turn starts with an empty <think>\n\n</think>\n\n block). The model speaks in the assistant role, and to have it open a conversation you start with an empty user turn. Use the tokenizer from caseyeasd/claudefriend, whose chat template defaults to that format.

Serving with vLLM:

vllm serve Qwen/Qwen3.5-9B-Base --enable-lora --max-lora-rank 32 \
  --lora-modules claudefriend=caseyeasd/claudefriend-lora

The base model's config ends generation at <|endoftext|>, but the model ends its turns with <|im_end|> (id 248046), so pass that as a stop token on every request, or it will keep going and write the next turn itself:

client.chat.completions.create(model="claudefriend", messages=messages, max_tokens=160,
                               extra_body={"stop_token_ids": [248046]})

With PEFT:

import torch
from peft import PeftModel
from transformers import AutoModelForImageTextToText, AutoTokenizer

tok = AutoTokenizer.from_pretrained("caseyeasd/claudefriend")
base = AutoModelForImageTextToText.from_pretrained("Qwen/Qwen3.5-9B-Base", dtype=torch.bfloat16)
model = PeftModel.from_pretrained(base, "caseyeasd/claudefriend-lora")
# when generating, stop at the end of the turn: model.generate(..., eos_token_id=[248046, 248044])