wchai/flint-frontier-cs-rollouts

Flint agent rollouts on Frontier-CS

This dataset holds every episode that Recursive Flint played in our long-horizon Frontier-CS game from October 24 to November 7, 2026: 9,649 episodes generated while training Flint with GRPO, and 3,114 episodes from evaluations of the original model and of RL checkpoints. Each row keeps the whole episode: every turn's reasoning, shell command and command output, every submitted program with its provisional and final judge scores, and the episode's final score. We release it for supervised fine-tuning and for the study of agent behaviour during RL.

The game

One agent, or a team of four, works on one Frontier-CS algorithmic problem in a git repository with a Linux shell. Each turn the model reasons and sends one bash command, and the harness returns its output. The command submit sends solution.cpp to the judge. The judge runs the problem's checker on the provisional test cases (files 1.in, 3.in, ...) and shows that score to the agent; it also scores the final cases (2.in, 4.in, ...), which the agent never sees. A problem with a single test case uses it for both (shared_case). The episode goes on until the agent's budget of generated tokens is used up. Its score S is the final score of the submission with the best provisional score (the earliest on a tie; 0 without a submission).

Two settings differ between RL and evaluation (see settings): RL episodes have 128k generated tokens, a turn limit of 16,384 tokens, a context of 110,000 characters and a 60-minute wall cap; evaluations have a turn limit of 32,768 tokens, a context of 400,000 characters and no wall cap. A turn that reaches the turn limit is cut (cut, finish reason length) and yields no command. When the chat outgrows its context, its oldest steps are dropped (dropped). Since October 31 the RL runs do not judge a resubmission of a program they already submitted (-dedup in harness), and since November 5 this ignores comments and layout. The harness tag (harness) names the game's version.

Configs

Config	Rows	What a row is
rl_rollouts	9,649	An episode sampled during GRPO training on the training problems
eval_episodes	3,114	An episode of an evaluation run (original Flint or an RL checkpoint)
eval_results	3,265	The evaluation driver's row for an episode: its score and metrics
problems	107	A problem, its split and the problem text the agent receives

Splits

problem_split is train67 (67 problems) or val40 (40 problems), our split of Frontier-CS v1.1's algorithmic problems. The RL runs on many problems trained on train67 problems only, mostly the 34 of in_signal34. The earlier single-problem runs (run20 and its tests, 280 rows of rl_rollouts) trained on val40's problem 0 (packing). val40 is our held-out set: leave its rows out of any training whose model you evaluate on val40.

Columns of rl_rollouts and eval_episodes

Column	Meaning
id	A stable id of the row
source	rl_rollout or eval
run	The RL run or the evaluation directory
rollout	The RL step that sampled the episode (-1 for evaluations)
trained_on	Whether the RL step trained on this episode (see below)
file, line	The source trace file within the run, and the line
pid	The Frontier-CS problem id
problem_split, in_signal34	The problem's split, and whether it is one of the 34 RL training problems
n_agents	1, or 4 for a team
budget	Generated tokens per agent
harness	The game's version tag
final_score	The episode's score S in [0, 1]
n_submissions, best_provisional, best_final	Judged submissions, and their best provisional and best final scores
n_turns, generated_tokens	Turns and generated tokens over all agents
model, base_model	Which weights played: the RL run's policy at that step, an RL checkpoint, or the original weights
settings	JSON: the game's settings, and for RL the problems and episodes per step, the seed and synchronous or asynchronous rollouts
code_commits	JSON: the harness commits that played the episode
system_prompt, first_user_messages	The chat's system message and each agent's first user message, rebuilt from the harness's prompt templates at the run's commit
submissions	JSON: every submission (agent, step, round, file, status, code, prov, final, infra), the end-of-episode auto-judged solution, and the best submission
trace	JSON: the whole episode: S, n, budget, harness, label, rounds, submissions, and per agent its turns (step, round, text with the reasoning, command, output, returncode, cut, finish_reason, format_error, completion_tokens, prompt_tokens, dropped)

In an asynchronous RL run (settings.rollouts names its staleness bound s), an episode's tokens may come from weights up to s steps older than its rollout. A few RL rollouts were sampled twice: a run stopped before the rollout's training step, and its resume sampled the rollout again with the same weights. trained_on is true for the attempt the run trained on and false for the earlier one (288 episodes); it is empty for evaluations.

Cleaning

Local paths, user, group and host names, private addresses and e-mail addresses are replaced by <work>, <user>, <group>, <host>, <ip> and <email>. The output of commands that show the host (ps, env, mount, df, id, /proc and similar) is replaced by <redacted: host information>, with output_redacted set on the turn. The per-turn server URL is removed. A scan of every string for these names and for API-key patterns found no remaining match.

Sources and license

The problems are Frontier-CS v1.1's (MIT license, FrontierCS Team). The model is Recursive Flint (Apache-2.0). This dataset is released under the MIT license.