Qwen3.5-9B — iLeen distilled base (merged 16-bit, alpha 16)
qwen35_9b_sys_2ep_a16.zip (18GB) — the base model for the gold SFT + DPO stage.
Merged 16-bit safetensors: unsloth/Qwen3.5-9B distilled on da'wah livechat data
(2 epochs, LoRA delta applied at alpha 16 = half strength, already baked into these weights).
Pairs with the adapters + data in iawos9/tempp.
Use
hf download iawos9/temp2 qwen35_9b_sys_2ep_a16.zip --local-dir .
unzip qwen35_9b_sys_2ep_a16.zip # needs ~36GB free (zip + unzipped)
mv qwen35_9b_sys_2ep_a16 base_qwen35_9b_sys_2ep_a16
Place it next to Qwen3_5_(9B)_DPO_no_unsloth.ipynb from the tempp bundle.
Contents: 4 safetensors shards + config.json, generation_config.json,
model.safetensors.index.json, tokenizer.json, tokenizer_config.json,
processor_config.json, chat_template.jinja.
Serving note: generation_config.json sets eos_token_id to <|endoftext|> (248044) while
the chat template ends turns with <|im_end|> (248046). Patch it before serving, or generation
runs past the turn boundary and invents user turns.