gpu-train-debug

An RL environment and dataset that trains and evaluates models on a real, scarce skill: diagnosing why an LLM training run failed or misbehaved, and naming the minimal correct fix.

The agent gets a training config, profiler numbers, and a raw log excerpt. It must output one diagnosis code and the corrected config fields as JSON. Grading is fully deterministic. There is no LLM judge and no human grader: the gold fix is computed by the generator, and the grader checks the final answer exactly.

Built in the Prime Intellect Verifiers format. The environment is published on the Prime Intellect Environments Hub as saloni-garg/gpu-train-debug (v0.4.1, public, MIT), and can be installed with prime env install saloni-garg/gpu-train-debug.

Splits

Frozen splits (seed 2027, deduplicated):

Split	Tasks	Purpose
train	1200	Training tasks across diagnosis, computed value, and two fault archetypes
hard	200	Interacting constraint tasks where a single signal is not enough
test	236	Held out evaluation tasks
ood	241	Out of distribution probe tasks with fresh combinations

Fields

• id: unique task id, for example train-two_lora_vocab-013
• archetype: the generator archetype that produced the task
• split: train, hard, test, or ood
• prompt: the full task prompt, including the diagnosis codes, the shared runbook, the config, profiler numbers, and the log excerpt
• gold: JSON string with the computed gold answer, with diagnosis and fix fields
• meta: JSON string with archetype metadata, tier, interaction, depth, and the runbook rules used

Note: gold and meta are stored as JSON strings because gold fix values vary in type across archetypes (booleans, integers, strings) and a diagnosis can be one code or a pair of codes for two fault tasks.

Why this exists

Labs are buying ML specific environments, and GPU training failures are a domain where generic synthetic data is weak: the evidence is numeric and cross referenced (config vs profiler vs log sequence). Every task is solvable from information inside the prompt, but not by keyword matching. All data is synthetic. No employer data, products, or internal numbers are included.

Validation results

Full validation was run on 2026-10-20 (Runpod A100). Supervised fine tuning (2 epochs) reached 0.83 on the hard split and 0.86 on the OOD split, versus 0.31 and 0.34 for the base model (Qwen2.5-3B-Instruct). Verifier based RL added no detectable gain beyond supervised fine tuning on the hard split, and a modest but measurable gain on the OOD split. See the accompanying paper, "Measuring the Marginal Value of Verifier-Based RL Beyond Supervised Fine-Tuning", for the full method, confidence intervals, and threats to validity.

License

MIT.