sweagent/iter2-rl-rollouts

iter-2 RL rollouts (combo_fb, Qwen3.5-35B-A3B)

Complete rollout + reward record for the iter-2 GRPO run: every trajectory the policy generated during training, with its graded reward. Preserved so the run stays re-analysable after its torch_dist checkpoints were retired (only latest-3 survive; the 11 HF milestones at iter_0/4/9/.../44/49 are the durable checkpoint record).

Run

base model	Qwen3.5-35B-A3B
init	iter_49 of the iter-1 RL run (hf_milestones_diffrecon/iter_0000049_hf)
harness	combo_fb (contract-ground + git-add-N + wall-clock valve + feedback)
algorithm	GRPO with dynamic sampling, kl-loss-coef 0.00, lr 1e-6 constant
batch	rollout-batch 32 x n-samples 4 (max 8), GBS 128
steps	0-49 (50 steps completed of a configured --num-rollout 300)
task pool	combined_0630_clean.jsonl — 2,484 tasks (SWE-rebench V1/V2 + Scale-SWE)
grading	azure-modal sandbox, F2P/P2P; resolved = full pass

Training was paused at ckpt 49, not run to completion — it stopped for checkpoint evaluation, and --load == --save means it can resume losslessly from iter_49.

Contents

Training and eval rollouts are in SEPARATE archives. The training set is the input to iter-3's hard-task search, so it is usable without filtering.

file	items	what
trajectories_00..05.tar.gz	23,139	training trajectories: messages, model_patch, exit_status, token_ids, loss_mask
rewards.tar.gz	20,248	training grading: resolved, tests_run/passed, f2p/p2p, error
eval_rollouts.tar.gz	5,330	in-training eval only (swe_val 200 + swe_multi 100), traj + reward

3.13 GB gzipped. Files are named <instance_id>_<sample_index>.json (trajectory) and <instance_id>_<sample_index>_rewards.json (reward), so the two join on name.

Training-set census

rollouts	20,248
unique tasks touched	2,417 of 2,484 (67 never sampled)
resolved	10,265 → 50.7%
rollouts per touched task	~8.4
per step	~405 (vs the 128 configured floor — dynamic-sampling oversampling)

Notes for re-analysis

• Training and eval were mixed on disk. Both land in the same trajectory directory, so a naive count gives 22,774 rollouts over 2,710 unique tasks — more tasks than the 2,484-task pool contains, which is the tell. The split here is by membership in the three id lists (combined_0630_clean, swebench_200_random, swe_multi_100); it accounts for every file with none unclassified. Eval resolves at ~70% against training's 50.7%, so pooling them biases both.
• Not every trajectory has a reward. 23,139 trajectories vs 20,248 rewards — a gap of 2,891. Rollouts that produce no patch (LimitsExceeded / timeout) are never sent to the sandbox, so they have no reward file. They count as unresolved. This is not data loss.
• never-solved is not never-attempted. 67 of the 2,484 tasks were never sampled at all. A task with 0 rollouts carries no difficulty signal; a task with 8 rollouts and 0 resolves does. Keep them separate when building a hard-task set.
• Two source directories. The run restarted once after an OOM, so rollouts are split across 20260814_060031_* and 20260815_091704_*. Any total must sum both — reading only the newer one silently drops roughly the first day.
• Per-step binning: these files carry no step field. Bin by file mtime against the training log's rollout N: markers, assigning a sample to the first marker at/after its mtime.
• No group_info in this run (the sibling combo2-rl-rollouts has 2,038 records).

Held-out result

SWE-V 500, combo_fb harness, avg@3, temp 0.95 / top_p 0.95 / top_k 20:

checkpoint	mean	sd	vs iter-1 endpoint (68.8)
iter_34	68.9	0.8	+0.1
iter_39	68.9	0.7	+0.1
iter_44	70.2	0.7	+1.4
iter_49	70.3	1.5	+1.5

Related

• sweagent/combo2-rl-rollouts — the equivalent record for the previous run (note: that one is unfiltered, mixing training and eval rollouts together)
• sweagent/coevolve-swev-grid-trajs — held-out SWE-V evaluation grid