diffrecon-rl-iter34

Co-evolve iter-1 RL checkpoint iter_34 (Qwen3.5-35B-A3B).

• Run: iter-1 RL (GRPO + dynamic sampling), harness = diff-reconcile, init = vanilla Qwen3.5-35B-A3B
• This is checkpoint 34 of 50 from that run.

Held-out SWE-bench Verified (/500, temp/top_p 0.95, 3-pass)

harness	score	n
diff-reconcile	67.9 +/- 1.5	5
combo_fb	68.1 +/- 1.7	3

For reference, iter_49 from the same run: 68.2 +/- 0.8 (diff-reconcile, n=6), 68.8 +/- 1.2 (combo_fb, n=10 pooled). The checkpoints are statistically tied on SWE-V.

In-training eval at this step

swe_val 65.3% (128/196), swe_multi 63.4% (single-pass proxy sets).

Not to be confused with

sweagent/rl-combo-iter34 -- that is combo2's iter_34, a different run. This repo is the iter-1 diff-reconcile RL run.