File size: 3,887 Bytes
63f41a6 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 | {
"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"
]
]
}
|