sashavoitovych2001/qwen2.5-3b-holmes-no-robots-lora-20260916

Qwen2.5-3B Holmes — LoRA adapter

Research coursework release from CS2881R Assignment 1. Model: no_robots_lora_seed_20260916, final checkpoint 980.

Trained from Qwen/Qwen2.5-3B-Instruct on 7,836 single-turn No Robots examples with unchanged prompts and assistant responses rewritten in Holmes's voice by gpt-5.6-luna. Development: 869 examples. No raw-canon examples are mixed into this model.

Training: two epochs, seed 20260916, learning rate 2e-05, maximum sequence length 1,024, effective batch size 16. Assistant tokens only contribute to chat loss.

LoRA rank 8, alpha 16, dropout 0.05; q/k/v/o attention projections. This is an unmerged adapter requiring the base model.

Selection and evaluation

Selected within its method by the highest equal-weight mean of GSM8K and No Robots persona scores across six final candidates. Ties use the earliest seed within a condition; full SFT seeds 20260916 and 20260917 tie. See selection.json and judge_comparison.csv for the full comparison.

Frozen v3 gpt-5.6-luna judge; scale 1–3 (1 = no recognizable persona). Each dataset has 60 prompts per model. Greedy generation, 512-token cap, default Qwen chat template, no explicit Holmes instruction.

Equal-weight mean: 1.103788. Holmes reply-token perplexity: 33.5136; GSM8K control perplexity: 2.5239 (192 examples each).

These scores are close to the scale floor, not evidence of a strong Holmes persona. Invalid judgments remain missing. GSM8K recovery stopped with errors; not all missing judgments were recovered. Small differences are descriptive, not statistically established. These sets were used for selection and are not independent post-selection tests. The No Robots LR sweep did not meet the original math-regression cap and used the documented fallback; math preservation is not established.

Usage

Install torch, transformers, accelerate, and peft. The repository ID below points to the published model.

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

repo_id = "sashavoitovych2001/qwen2.5-3b-holmes-no-robots-lora-20260916"
tokenizer = AutoTokenizer.from_pretrained(repo_id)
from peft import PeftModel
base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-3B-Instruct", torch_dtype=torch.bfloat16, device_map="auto")
model = PeftModel.from_pretrained(base, repo_id)
model.eval()
messages = [{"role": "user", "content": "How can I find a misplaced key?"}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
with torch.inference_mode():
    output = model.generate(**inputs, max_new_tokens=512, do_sample=False)
print(tokenizer.decode(output[0, inputs.input_ids.shape[1]:], skip_special_tokens=True))

License and provenance

The base model uses the Qwen Research License, preserved in LICENSE. The No Robots dataset is CC BY-NC 4.0. This release is for non-commercial research/evaluation. See NOTICE for attribution and modifications. Optimizer states and raw paid API attempts are excluded.