Qwen2.5-7B-Instruct converge-collective-v1

The first cross-family merge produced entirely in converge — CRDT-based, convergence-guaranteed distributed model merging. Same 9-model, 4-family cohort as Optitransfer/Qwen2.5-7B-Instruct-borg-merge-v1, re-executed under converge's exact-integer arithmetic and certified by a convergence digest.

The project

converge is building the first model nobody owns: a living collective that absorbs signed fine-tune contributions forever. Two mathematical guarantees: every node converges byte-identically (CRDT + an associative integer merge monoid), and every merge is provable (signed receipts, offline-verifiable). The people's tier is free forever — every contribution is cryptographically attributed, knowledge only compounds, and the model architecturally cannot forget (contributions are generational capital: each new anchor generation re-assimilates the full archive). Commercial tiers sell guarantees AROUND the collective — never pieces OF it (the Two-Tier Charter).

This merge (measured, paired, same-settings)

9 models, 4 architecture families (llama_family, phi, neox, opt), anchor Qwen2.5-7B-Instruct + 8 donors
Kernel: the proven E3 anchor merge — per-role relative tolerance clipping (attn 0.05 / FFN 0.20), rank-3 SVD consensus, Procrustes alignment — computed through the int24 pre-committed-scale integer monoid and stored as int16 lattice shards (wire grid = merge grid)
Integrity gates: embeddings / lm_head / all norms pass through the anchor bit-exact (cross-tokenizer merging is meaningless, not attempted)
Telemetry: 280 keys merged / 59 gated passthrough, 808 alignments, 112 consensus-filtered keys

Convergence certificate

Digest (SHA-256 over the int16 lattice grid): 178520f17063358cd23c78dbf7bf1e6b2d61ed5bf94d6befe040c63bcfde3025
Reproduce the merged state: any node folding the same inputs through the same contract derives the identical digest (test-proven under donor permutation; order/replica-invariant by construction).
Receipt: 0 signed contributions, issuer ff56dccfcca9ecf0... — receipt.json ships with the model; verify offline with the converge receipts layer (re_execute re-derives the merge and compares hashes).

Benchmark results (paired anchor vs merged, identical settings)

Evaluated with lm-eval: bf16, batch 4, seed 42,1234,1234, same harness for both models. Absolute scores are NOT comparable to other publications' setups — only the paired deltas are claims.

Benchmark (metric)	Anchor	Merged	Delta	Verdict
arc_challenge	acc,none	0.6442	0.6391	-0.0051
arc_challenge	acc_norm,none	0.6698	0.6681	-0.0017
gsm8k	exact_match,flexible-extract	0.8264	0.8294	+0.0030
gsm8k	exact_match,strict-match	0.7627	0.8059	+0.0432
hellaswag	acc,none	0.6223	0.6066	-0.0157
hellaswag	acc_norm,none	0.8138	0.7932	-0.0206
piqa	acc,none	0.7943	0.7922	-0.0022
piqa	acc_norm,none	0.8036	0.7992	-0.0044
truthfulqa_mc2	acc,none	0.6463	0.6255	-0.0208
winogrande	acc,none	0.7127	0.7174	+0.0047

Honest trade profile: a static merge is a trade, not a lift — axes that regress are visible above and are exactly what the ratchet gate (monotone-by-gate assimilation) protects against in the living collective.

Method (five steps, converge-native)

Canonicalize all 9 models into the shared key namespace (per-family detectors; round-trip verified)
Zero-pad bridge to the anchor's shapes; Procrustes-rotate attention tensors into the anchor basis (SVD direction verified)
Rank-3 SVD consensus across donors (content-ordered — a pure function of the donor set)
Clip donor deltas RELATIVE to the anchor per role; average through the associative integer monoid (pre-committed scale)
Quantize to int16 lattice shards at per-key cohort scales; assemble into the anchor's native layout

Limitations (on record)

Single cohort (the bootstrap 9); population assimilation with the ratchet gate is the next phase
Honest axes of regression are shown, not hidden (see table)
Cross-family basis alignment beyond zero-pad+Procrustes remains open research (activation-matched alignment)