Jerry9999/RSIChallengeCMU

RSIChallenge: sandbox instances and agent episodes

Training and evaluation environments built from real PostTrainBench runs, and the episodes of agents run on them. Uploaded 2026-11-12 for review of instance quality.

What an instance is. A real moment of a real PostTrainBench run: an agent was post-training a small base model (Qwen3-1.7B-Base, Qwen3-4B-Base, SmolLM3-3B-Base or gemma-3-4b-pt) for one benchmark, and the instance is the directory it left behind at that moment. A new agent gets the same prompt, container, GPU and official grader, a short budget (minutes of tool time and a cap on replies), and is told only that an earlier session left the directory. Nothing says what is wrong.

• End-of-run instances (instances/end_of_run/): a finished run whose final_model/ holds one fault (wrong decoding config, missing stop token, a fallback model, an interrupted copy, a lost tokenizer, ...), or a healthy control. Score: the benchmark's own evaluate.py on final_model/ at the end, normalised so that doing nothing scores 0 and the reference fix scores 1.
• FT instances (instances/ft/): the moment just before a training launch that crashed in a real run. Score: that same command is rerun on the workspace the agent leaves, and passes if it gets past the original crash point. No weights: an FT instance is code, data and a command; the base model is downloaded by its HF id.

Each instance directory holds state/ (what the agent sees) and meta.json (family, sub-type, benchmark, base model, source run, budget, no-op and oracle scores; the agent never sees it). FT instances also hold verify.json (the command and its time limit) and ref/ (the reference fix).

Contents

Path	What	Count
manifest/sandbox_instances.csv	every instance, frozen or not: set, kind, family, benchmark, split, base model, source run, no-op, oracle, budget, where its weights are	167 rows
instances/end_of_run/	end-of-run instances (92 frozen)	80 with weights, 12 not runnable as built
manifest/hub_instances.csv	per end-of-run instance: what is on the Hub, layout, runnable or not, and why	92 rows
manifest/left_out_files.tsv	files not uploaded, with the reason	15
instances/ft/	frozen FT instances	44
manifest/ft_dropped.json, manifest/ft_index_all75.json	the 31 FT candidates that failed a gate, with the reason	—
episodes/index.csv	one row per episode: model, scaffold, instance, repeat, benchmark, split, official accuracy, S (unclipped), FT replay status, tool time, replies, how it ended, and which archive holds it	1,265 rows
episodes/<scaffold>_<model>_0h_offline_sbx.tar.gz	all episodes of one agent; inside, <benchmark>_<base model>_<instance>[m30]r<repeat>/	14 archives

Status (2026-11-12): complete. 92 frozen end-of-run instances, 44 FT instances, 1,265 episodes. manifest/hub_instances.csv gives, per end-of-run instance, what is here and whether it can be run as built:

• 80 runnable: final_model/ with weights, plus the agent's other checkpoints and data.
• 12 not runnable as built: their faulted final_model was never materialised, because they were built from public-corpus runs, which have traces but no weights. Their code, logs and data are here for review. (par-gs-b4-glm53fl-l1a1 has an intermediate checkpoint, state/ckpt_run2/, but no final_model.)
• 15 files left out for holding credential-like strings, listed in manifest/left_out_files.tsv.

Two layouts. Instances built on babel (F1 and part of v2) are raw files under state/. Instances built or kept on the aws cluster (sets "v2 built on aws", v2b, v2c) hold state_small_files.tar.gz (every file under 1 MB of state/) plus the larger files as raw files under state/; to rebuild such an instance, extract the archive into the instance directory and keep the raw files beside it: tar -xzf state_small_files.tar.gz.

Splits. Instances are built from runs on the PostTrainBench dev benchmarks (gsm8k, bfcl, aime2025, gpqamain) and, for 29 end-of-run instances, humaneval. humaneval, arenahardwriting and healthbench are the held-out benchmarks of the project: the split column marks them, and those instances are for evaluation and review only, never for training.

An episode

Episodes are packed one archive per agent (the Hub limits API calls, and 22k small files would exceed it). m30 in a directory name marks an F1 instance run with a 30-minute budget instead of 10. S_unclipped in the index is (accuracy − no-op) / (oracle − no-op), before clipping to [0, 1].

File	What
prompt.txt	the prompt the agent got
solve_out.txt.gz	the full trajectory: Claude Code stream-json (every assistant turn, tool call and tool result) or Codex json events
clock.json, time_taken.txt	tool time charged, wall time
gateway_requests.jsonl.gz	per-call metadata of the model gateway (tokens, latency, status; no message content)
final_eval*, metrics.json	end-of-run: the official evaluation of the model the agent left
replay.json	FT: the rerun of the named command and whether it passed
final_model/	the config files of the model the agent left (weights and tokenizer copies removed)
run_info.txt, output.log, error.log	harness logs

Agents: Claude Code with claude-opus-5, claude-opus-4-8, claude-opus-4-7, claude-opus-4-6, claude-opus-4-5, claude-sonnet-4-6; Codex with gpt-5.6-sol and gpt-5.5; Claude Code with five self-hosted open models (gemma-4-31b, qwen3.5-27b, qwen3.6-27b, qwen3.6-35b, qwen3.8-27b) on the first 16 instances only. _reverify_controls/ holds re-verifications of control scores (episodes/_reverify_controls.tar.gz).

Notes

• Redaction. Files were scanned for API keys, tokens and internal hostnames before upload. One per-run gateway key (an ephemeral key, not a provider key) printed in one trajectory was replaced by <redacted>. The remaining matches are public examples inside training data: AWS's documented AKIAIOSFODNN7EXAMPLE and a placeholder GitHub token in a BFCL function schema (the same string as in the public Salesforce/xlam-function-calling-60k data). Fifteen files whose credential-like strings could not be traced to a public source were left out (manifest/left_out_files.tsv).
• Licenses. Trajectory data of the public PostTrainBench corpus is Apache-2.0. Weights fine-tuned from gemma-3-4b-pt are subject to the Gemma Terms of Use; Qwen3 and SmolLM3 derivatives are Apache-2.0. Benchmark test sets copied into run directories keep their own licenses.
• Provenance. Built by the RSIChallenge project; the design is documented in its repository (docs/plans/sandbox_task_design.md, docs/plans/objective.md), and the instance list is regenerated by experiments/sandbox/instance_manifest.py.