| { |
| "columns": [ |
| "Stage", |
| "Leanstral 1.5 (reported)", |
| "Our 8B / 4x A10G analogue", |
| "Data (see Datasets)", |
| "Gate to advance", |
| "Box-hours", |
| "Source", |
| "Confidence", |
| "Status" |
| ], |
| "widths": [ |
| 16, |
| 46, |
| 46, |
| 34, |
| 34, |
| 12, |
| 40, |
| 10, |
| 14 |
| ], |
| "rows": [ |
| [ |
| "Base", |
| "Mistral Small 4: 119B MoE, 128 routed + 1 shared expert, top-4, ~6B active, 36 layers, MLA", |
| "Goedel-Prover-V2-8B (dense Qwen3-8B derivative) - already Lean-trained", |
| "-", |
| "Reproduce published baseline", |
| "~12", |
| "interim-report-leanstral-base-midtraining-and-compute-provenance", |
| "High", |
| "Done (82.8% pass@32)" |
| ], |
| [ |
| "Mid-training", |
| "6.5B dedup Lean tokens + code-agent / instruction data; ratios unreported", |
| "Skipped: base is already Lean-specialised; 6.5B tokens ~24 days here", |
| "Datasets: S1", |
| "-", |
| "~575", |
| "interim-report-leanstral-base-midtraining-and-compute-provenance", |
| "High", |
| "Out of scope" |
| ], |
| [ |
| "SFT", |
| "50% Lean code-agent traces, failure-mode filtered; teachers unreported", |
| "Iterated verifier-filtered self-distillation (LoRA) on band problems + compiler-error correction pairs", |
| "Datasets: S2a, S2b; own pool rounds", |
| "valid pass@8 gain >= noise; no regression", |
| "~55 per round", |
| "interim-report-data-vs-rl-lever-and-8b-4xa10g-reproducibility", |
| "Med", |
| "Queued" |
| ], |
| [ |
| "Prompt filter", |
| "Keep prompts with low-but-nonzero SFT solve rate (8 rollouts)", |
| "Band (0, 0.75] from pool round n=8; ~11% of pool", |
| "Pool round 0", |
| "-", |
| "~25", |
| "in-this-work-we-present-leanstral-a-series-of-generalist-code-agent-models-for", |
| "High", |
| "Running" |
| ], |
| [ |
| "RL", |
| "CISPO, loss -(1/sum|o|) sum sg(min(pi/pi_sample, c)) A log pi; c unreported; outcome broadcast to all tokens; tool/prompt tokens masked; mix 50% LeanGym-PR / 20% single theorem / 30% prove-or-disprove multi-turn", |
| "Small on-policy LoRA CISPO, c~5, no lower clamp, neg-adv low-ratio sequence mask, 1-2 updates/rollout, single-turn + correction mode, <=40k ctx", |
| "Datasets: S3", |
| "paired valid gain; collapse tripwires", |
| "~15", |
| "interim-report-objective-choice-cispo-vs-grpo-effect-size-controlled", |
| "Med", |
| "Planned" |
| ], |
| [ |
| "Reward / verifier", |
| "Format partial + compile with standard axioms only (#print axioms); native_decide banned; LeanInteract + modified SafeVerify", |
| "Strict verifier (P1) + red-team fixes + OS sandbox; same verifier both arms", |
| "Datasets: S4", |
| "0 false accepts on adversarial suite", |
| "-", |
| "in-this-work-we-present-leanstral-a-series-of-generalist-code-agent-models-for", |
| "High", |
| "Running" |
| ], |
| [ |
| "Averaging", |
| "Not in Leanstral report (Goedel-V2 item)", |
| "alpha 0.6-0.9 between SFT round and RL/next round, chosen on valid", |
| "-", |
| "valid pass@32 restored", |
| "~6", |
| "interim-report-rl-adds-vs-sharpens-on-lean-passk", |
| "Med", |
| "Planned" |
| ], |
| [ |
| "FP32 LM head", |
| "Not in Leanstral report (MiniMax-M1 item)", |
| "Kept (cheap); expected ~0 gain on dense 8B", |
| "-", |
| "-", |
| "0", |
| "interim-report-objective-choice-cispo-vs-grpo-effect-size-controlled", |
| "Med", |
| "Done (implemented)" |
| ], |
| [ |
| "Scaffold", |
| "Mistral Vibe agent; compaction every 200k by mistral-small; persistent files/lemmas", |
| "Single-turn whole proof + bounded correction turns; 200k windows infeasible (~29 GB KV)", |
| "-", |
| "-", |
| "-", |
| "interim-report-multiturn-compiler-feedback-vs-singleturn-and-sampling", |
| "Med", |
| "Planned" |
| ], |
| [ |
| "Bug finding", |
| "Aeneas Rust->Lean; generate properties; pass@4 prove + pass@4 negation; 30 turns; 57 repos -> 47 flagged -> 11 real -> 5 new", |
| "Same pipeline on a small crate set after the prover is final", |
| "Datasets: S5", |
| "Manually confirmed bugs", |
| "-", |
| "in-this-work-we-present-leanstral-a-series-of-generalist-code-agent-models-for", |
| "High", |
| "Planned" |
| ] |
| ] |
| } |
|
|
