RLTT/Affine-sn120-cand4-checkpoints

Affine SN120 round-4 checkpoint sweep (LoRA adapters)

Every 50-step checkpoint from a calibration fine-tune of the Affine SN120 reign-12 king (d76150805915a988723d16aed4c45dc8e61bedeb08fa8599289285792f3f7795, published as affine-14-ed-1).

Adapters only. To reconstruct a checkpoint, merge one into the base:

from transformers import AutoModelForCausalLM
from peft import PeftModel
m = AutoModelForCausalLM.from_pretrained(BASE, dtype="bfloat16")
m = PeftModel.from_pretrained(m, "checkpoint-150")
m = m.merge_and_unload()

The base is public at https://models.affine.io/models/sha256/d76150805915a988.../.

What the run optimises

Live score is turn = min(R, G); R >= 0 by Jensen, so a turn only goes negative through the grounding leg G = w - |m - mu| or through a forfeit (-0.1, for no parseable action or a reply that never closes </think>).

In our real duel (chal-00474, +0.000460 margin vs the reign-12 king) the entire deficit was forfeits: 4.67% of turns, concentrated in text (parse 0.817) and boxed (parse 0.731). Round 4 targets that directly.

Training

• Targets generated locally from the frozen teacher Qwen/Qwen3.8-27B: 8 samples per turn, keeping the one whose lpC(z|x) is closest to the mean of the 8 — the teacher's most typical thought, not its most likely one (that lands above the band and scores negative).
• Median centring deviation 0.0038, vs 0.0084–0.0116 for a best-of-3 pick from published duel references.
• Hard caps on the thought span (text ≤ 300 tokens, others ≤ 400), well under the 1792-token contract budget, because </think> failing to arrive is what produces the forfeit.
• Corpus epoch 31, all five live dialects including terminus_json.
• LoRA r=32 on attention (both full-attn and GDN linear-attn layers) plus the shared expert; routed experts untouched.

Caveat

Advisory SWE-rebench benches were not run. The subnet's own research records a negative rank correlation between duel margin and SWE-rebench on the live board, so a min(R, G) gain should not be read as a coding capability gain.