Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,53 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: mit
|
| 3 |
+
tags:
|
| 4 |
+
- swe-bench
|
| 5 |
+
- reinforcement-learning
|
| 6 |
+
- agentic
|
| 7 |
+
- rollouts
|
| 8 |
+
---
|
| 9 |
+
|
| 10 |
+
# combo2 RL rollouts — Qwen3.5-35B-A3B
|
| 11 |
+
|
| 12 |
+
Complete rollout + reward record for the **combo2** GRPO run: every trajectory the policy
|
| 13 |
+
generated during training, with its graded reward. Preserved so the run stays re-analysable
|
| 14 |
+
after its torch_dist checkpoints were retired.
|
| 15 |
+
|
| 16 |
+
## Run
|
| 17 |
+
|
| 18 |
+
| | |
|
| 19 |
+
|---|---|
|
| 20 |
+
| base model | Qwen3.5-35B-A3B |
|
| 21 |
+
| init | `sweagent/practical-diffrecon-ep2` (the iter-1 RFT ckpt) |
|
| 22 |
+
| harness | combo (contract-ground + git-add-N + wall-clock valve) |
|
| 23 |
+
| algorithm | GRPO with dynamic sampling, `kl-loss-coef 0.00`, lr 1e-6 constant |
|
| 24 |
+
| batch | rollout-batch 32 x n-samples 4, GBS 128 |
|
| 25 |
+
| steps | 0-54 |
|
| 26 |
+
| task pool | `combined_0630` (SWE-rebench V1/V2 + Scale-SWE) |
|
| 27 |
+
| grading | azure-modal sandbox, F2P/P2P; `resolved` = full pass |
|
| 28 |
+
|
| 29 |
+
**Superseded by** the base -> iter_49 -> iter-2 line; published as a record, not as the current method.
|
| 30 |
+
|
| 31 |
+
## Contents
|
| 32 |
+
|
| 33 |
+
| file | items | what |
|
| 34 |
+
|---|---|---|
|
| 35 |
+
| `trajectories_00..05.tar.gz` | 27,787 | full trajectories: messages, `model_patch`, `exit_status`, `n_steps`, prompt/response token_ids, loss_mask |
|
| 36 |
+
| `rewards.tar.gz` | 26,610 | per-sample grading: `resolved`, `resolution`, `tests_run/passed`, f2p/p2p rates, `error` |
|
| 37 |
+
| `group_info.tar.gz` | 2,038 | per-instance group records (`saved_at`, `reward`, `sample_index`, `token_length`) |
|
| 38 |
+
|
| 39 |
+
34.8 GiB raw, ~2.8 GB gzipped. Files are named `<instance_id>_<sample_index>.json`
|
| 40 |
+
(trajectory) and `<instance_id>_<sample_index>_rewards.json` (reward), so the two join on
|
| 41 |
+
name.
|
| 42 |
+
|
| 43 |
+
## Notes for re-analysis
|
| 44 |
+
|
| 45 |
+
- **Not every trajectory has a reward.** Rollouts with no patch (`exit_status` of
|
| 46 |
+
`LimitsExceeded` / `TimeExceeded`) are never graded — they count as unresolved. The
|
| 47 |
+
~1,200 gap between traj and reward counts is that, not data loss.
|
| 48 |
+
- **Per-step binning**: these files carry no step field. Bin by file mtime against the
|
| 49 |
+
training log's `rollout N:` markers, assigning a sample to the **first marker at/after**
|
| 50 |
+
its mtime — that line prints when a finished batch is handed to `actor_train`, so a
|
| 51 |
+
step's rollouts precede its marker.
|
| 52 |
+
- **`resolved` is the ground truth**, not the in-pipeline `resolve=N/N` stdout print, which
|
| 53 |
+
is known to read 0.
|