Result (worst-of-2 medal, 11 dev cells, bf16-vs-bf16)
Table with columns: Local-base, This model (r2), Teacher Qwen3.7-Plus | Local-base | This model (r2) | Teacher Qwen3.7-Plus |
|---|
| worst-of-2 | 78.0 | 85.5 | 78.5 |
Leads on the customer-honest (worst-seed) metric by being tight: near-best everywhere rather than
brilliant-here / broken-there. Gain is modest and of uncertain magnitude (small, saturated dev set) —
see the dataset card for the full
protocol, honesty guards, and the round-1 negative control.
Use
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3.6-27B", dtype="bfloat16", device_map="auto", trust_remote_code=True)
model = PeftModel.from_pretrained(base, "Zeeshank1994/factory-solver-qwen3.6-27b-rft-r2")
tok = AutoTokenizer.from_pretrained("Qwen/Qwen3.6-27B", trust_remote_code=True)
To serve merged (bf16, SGLang): merge the adapter shard-level into the full base checkpoint
(the ForConditionalGeneration wrapper must be preserved — a plain AutoModelForCausalLM merge strips
the wrapper and SGLang then refuses to serve it).