[VDR] TUT-f93062c5 · Offline-training filter drops 100% of real rollouts (requires a 'success' field Gym never emits)

Open Beginner friendly

#2,235 opened on Jul 30, 2026 1 comment 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

Assessment

Difficulty: 2/5
Estimated time: 1-3 hours
Newbie friendliness: 76/100

Issue type: Bug
Clarity: Mostly clear
Activity status: Active
Tech stack: python
Domain: documentation

Research direction

Inspect filter_rollouts() at fern/versions/latest/pages/training-tutorials/offline-training-w-rollouts.mdx:110 and compare its success check with example_multi_step/data/example_rollouts.jsonl. Run the tutorial against real Gym rollout rows, then verify filtered_rollouts.jsonl and the downstream SFT/DPO formatting outputs contain data instead of being empty.

Written by the indexing model from the issue text.

Description

vdr

filter_rollouts() ANDs in rollout.get('success', False). Gym rollout rows do not contain a success key — a shipped rollout row (example_multi_step/data/example_rollouts.jsonl) has keys {accuracy, order_instruction_following_failure, original_term_minefield_hit, parsed_synonym_values, response, responses_create_params, reward, set_overlap}. Running the tutorial's own code on real Gym rollouts writes an empty filtered_rollouts.jsonl, so every downstream step (SFT/DPO formatting) yields empty output with no error.

Suggested fix: Drop the success condition (reward already encodes it) or key it off a field Gym actually writes. Also note the page is marked experimental but still presents this as the first runnable step.

Details


Severity	🟠 high
Path	`fern/versions/latest/pages/training-tutorials/offline-training-w-rollouts.mdx:110`
Finding ID	`TUT-f93062c5`
Found by	`tutorial-walker`
Workstream	[VDR][v0.5.0] Tutorials fail when followed verbatim
Status	⟳ also reported 2026-06-30 (as `TUT-2ef29637`) — still unfixed

Evidence

TOP KEYS: ['accuracy','order_instruction_following_failure','original_term_minefield_hit','parsed_synonym_values','response','responses_create_params','reward','set_overlap'] ; has success: False | has task_id: False

_{DX review Gym-1-20260730170035 · target ref e446e4f4 · 2026-07-30 · tracker #2214}

Dominant language: Python
Stars: 1.2k
Forks: 349
Avg merge: 2d 40m
Merged PRs (30d): 320

Contributor guide

Open the contributing guide

First steps

Read the whole issue, then the project's contributing guide.
Comment on the issue to say you are picking it up — it saves two people doing the same work.
Fork the repository and make your change on a branch.
Open a pull request that references the issue number.

Similar issues

[BUG] GkeTpuCluster queries metadata.google.internal for MEGASCALE_SLICE_ID even when TPU_SKIP_MDS_QUERY is set Open

Difficulty 2/5 1-3 hours Newbie friendliness 88/100

jax-ml/jax#40854 · Opened Sep 22, 2026
[pcied] Honour PCIE_DETACH_INFO on platforms that are not smartswitches Open

Difficulty 2/5 1-3 hours Newbie friendliness 82/100

sonic-net/sonic-platform-daemons#900 · Opened Sep 21, 2026
MCP client has no httpx timeout, so the event stream fails after 5 seconds Open

Difficulty 2/5 1-3 hours Newbie friendliness 78/100

Azure/azure-functions-agents-runtime#228 · Opened Sep 21, 2026
feat(mcp): friendlier error when `run_cloud_sync` is called on a disabled connection Open

enhancement

Difficulty 2/5 1-3 hours Newbie friendliness 72/100

airbytehq/PyAirbyte#1205 · Opened Sep 21, 2026
pn7160: reader goes permanently deaf — post-read IRQ-low check fails reads that succeeded Open

Difficulty 2/5 1-3 hours Newbie friendliness 74/100

esphome/esphome#19452 · Opened Sep 21, 2026

More Python issues