--- license: apache-2.0 base_model: Qwen/Qwen3.5-2B-Base library_name: peft pipeline_tag: text-generation datasets: - allenai/tulu-3-sft-personas-instruction-following - databricks/databricks-dolly-15k tags: [lora, peft, qwen3.5, instruction-tuning] --- # instruct Instruction-tuned Qwen3.5-2B-Base: a rank-8 LoRA trained with memorize.sft_torch (the sft-run2 recipe) on the memorize.sft_data mix. Start point for the label-split experiment. From [Memorize](https://github.com/AriSpiesberger/Memorize), a small lab for poking at Qwen3.5-2B. ## use ```python import torch from peft import PeftModel from transformers import AutoModelForCausalLM, AutoTokenizer tok = AutoTokenizer.from_pretrained("Qwen/Qwen3.5-2B-Base") base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3.5-2B-Base", dtype=torch.bfloat16) model = PeftModel.from_pretrained(base, "Arisp/memorize-instruct") ids = tok.apply_chat_template( [{"role": "user", "content": "Name three fruits."}], add_generation_prompt=True, enable_thinking=False, return_tensors="pt", return_dict=True, )["input_ids"] out = model.generate(ids, max_new_tokens=100, eos_token_id=tok.convert_tokens_to_ids("<|im_end|>")) print(tok.decode(out[0, ids.shape[1]:], skip_special_tokens=True)) ``` Use the chat template with `enable_thinking=False` (how it was trained), and stop on `<|im_end|>`: the base model's own config only stops on `<|endoftext|>`. ## results | | base model | this adapter | | --- | --- | --- | | validation loss (200 held-out chats) | 1.361 | 1.250 | | instruction probes passed (greedy) | 15/19 | 18/19 | | MMLU-Pro, direct letter, n=500 | 37.4% | 38.0% | Probes are short prompts with checkable output constraints ([memorize/probes.py](https://github.com/AriSpiesberger/Memorize/blob/main/memorize/probes.py)). MMLU-Pro is scored from the letter right after `ANSWER:` with no reasoning, on label-split C, split seed 0 ([memorize/label_split.py](https://github.com/AriSpiesberger/Memorize/blob/main/memorize/label_split.py)). ## how it was made - **Data**: 1,800 constraint-following chats from tulu-3-sft-personas-instruction-following plus 600 Dolly-15k chats, 200 held out for validation, each at most 512 tokens; multiple-choice and step-by-step items dropped. Seed 0. - **LoRA**: rank 8, alpha 160 (mlx-lm's scale of 20), no dropout, on every linear layer of every block (attention, linear attention, MLP); 8.4M parameters. - **Training**: 1 epoch, 300 steps of 8 chats, Adam at lr 1e-5 (20 warmup steps from 1%, then cosine to 10%), gradient norm clipped at 20, bf16, loss on the assistant reply only. Seed 0. - **Environment**: Python 3.12.7, torch 2.11.0+cu128, transformers 5.18.0, peft 0.21.2, datasets 5.1.0, one RTX 4090 (Windows 11). `sft_torch.json` (in this repo) has every argument and the validation curve; `train.log` is the full training log. ## reproduce ```bash git clone https://github.com/AriSpiesberger/Memorize && cd Memorize git checkout aad35a9 # the code that trained it python -m venv .venv && .venv/bin/pip install -r requirements.txt # Windows: .venv\Scripts\pip python -m memorize.sft_data # data/sft/train.jsonl, valid.jsonl python -m memorize.sft_torch --out adapters/instruct ``` The data files should hash to sha256 `92234522d2e41621…` (train) and `fdcbf7bca4450523…` (valid) if the upstream datasets are unchanged.