Optitransfer: the converge programme

.gitattributes

@@ -0,0 +1,35 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,120 @@
+---
+title: Optitransfer
+emoji: 🔬
+colorFrom: indigo
+colorTo: blue
+sdk: static
+pinned: true
+tags:
+  - model-merging
+  - open-weights
+  - llm
+  - qwen2.5
+  - 7b
+  - text-generation
+  - reproducible-evaluation
+  - benchmarks
+  - leaderboards
+---
+
+<div align="center">
+
+# Optitransfer
+
+**Converge: building frontier-level open models from the work the open-source community has already published.**
+
+*One repeatable process at every model class. Every result is measured against its base model and published with per-item evidence.*
+
+[Models](https://huggingface.co/Optitransfer) · [crdt-merge](https://github.com/mgillr/crdt-merge) · [acfa-rs](https://github.com/mgillr/acfa-rs) · [Paper](https://arxiv.org/abs/2605.19373)
+
+</div>
+
+---
+
+## Purpose
+
+Frontier AI is concentrated in a handful of laboratories, because each generation of pre-training costs more than
+almost anyone else can spend. The open ecosystem has already paid that cost many times over. Thousands of fine-tuned
+models sit on public repositories, each one somebody's GPU-hours and each adding a skill to a shared base: mathematics,
+code, instruction following, a domain, a language.
+
+**Converge** is built to turn that work into frontier-level open models. It does not out-train the laboratories. It brings what
+the community has built into a single model for each model class, measures every step, keeps every gain and
+publishes every regression. The model is never finished: it improves release after release.
+
+## The goal
+
+**Reach frontier-level performance in every model class.** The programme starts with 7B, then moves to each larger
+size class and to other model families. At every class the model must improve on every axis that matters (reasoning,
+mathematics, code, instruction following and knowledge) without a material regression against its base. Each class
+is complete only when no public fine-tune improves any axis further.
+
+## The process
+
+The process is defined once and run the same way at every class, so every release can be rebuilt and re-measured.
+
+1. **Start** from a strong open base model in the class.
+2. **Discover** the compatible fine-tunes the community has published for it.
+3. **Assimilate** them into one model.
+4. **Measure** every axis against the base model on identical items, in the same session, on public benchmark protocols.
+5. **Release** a model only when it holds against its base and the previous release on the whole board, and publish every regression next to every gain, with per-item evidence. This is the release rule from the corrected releases on; v1 and v2 predate it, and their cards list their regressions.
+6. **Repeat** until nothing more improves, then carry the process to the next class.
+
+The construction method is proprietary. The evaluation is fully public: protocols, per-item outputs and the tools
+that recompute every number.
+
+## Where the programme stands
+
+| Release | Status |
+|---|---|
+| [converge v1](https://huggingface.co/Optitransfer/Qwen2.5-7B-Instruct-converge-collective-v1) | Public. The first 7B release, on Qwen2.5-7B-Instruct. It carries a merge defect that has since been corrected; see the notice on its card. |
+| [converge v2](https://huggingface.co/Optitransfer/Qwen2.5-7B-Instruct-converge-collective-v2) | Public. The current release, with a full validation board. It is built on v1, so it inherits the defect. |
+| corrected v1 and v2 | Built with the same recipes and the defect removed. Being benchmarked across the full board now. |
+| converge v3 | Next. Released only once it holds against its base model and the previous release on the whole board. |
+
+## How we report
+
+- **Paired, never selected.** Every score is the model minus its base on the same items, with a 95% confidence
+  interval and an exact significance test.
+- **Regressions are first-class.** Each card lists what went down as prominently as what went up.
+- **Reproducible only.** Public numbers come from reproducible runs on public protocols: HELM-protocol GSM8K and
+  MMLU, MMLU-Pro (TIGER-Lab), ZeroEval, EvalPlus HumanEval+ and MBPP+, MATH, AIME, ARC and IFEval.
+- **Corrected in public.** When a defect is found, the affected cards say so first.
+
+## Who it is for
+
+- **The open community.** A stronger open model at each class, with the evidence to check it.
+- **Healthcare.** A hospital network fine-tunes models on patient data that cannot leave its walls. Converge brings
+  those internal fine-tunes together inside its boundary, with a record of every change.
+- **Finance.** A regulated bank must show auditors how each model update was made, what changed and that nothing
+  regressed. Converge's per-item record is that audit trail.
+- **Sovereign AI.** A national programme needs frontier capability without depending on a few foreign vendors.
+  Converge grows by contribution, not by procurement.
+
+## How it is offered
+
+**A free model, with paid guarantees.** The public model stays open, and the free tier is never limited to sell a
+paid one. Around it, the first commercial product is auditable private aggregation: Optitransfer brings an
+organisation's own fine-tunes together inside its boundary, with the same measurement and evidence. Compliance
+evidence, sovereign deployments and private federations follow.
+
+## Open components
+
+- **[crdt-merge](https://github.com/mgillr/crdt-merge):** conflict-free (CRDT) merging for data, JSON, models and
+  agents, with the E4 trust-scoring engine ([paper](https://arxiv.org/abs/2605.19373)).
+- **[acfa-rs](https://github.com/mgillr/acfa-rs):** accountable federated aggregation with signed, offline-verifiable
+  receipts (Apache-2.0, [arXiv:2607.10305](https://arxiv.org/abs/2607.10305)).
+
+## Team
+
+**Ryan Gillespie**, founder, architect and operator · [@Optitransfer](https://huggingface.co/Optitransfer)
+
+- Fifteen years of verification and validation on safety-critical systems: railway signalling (Crossrail, ETCS),
+  medical devices taken to CE mark (Roche), and aviation (Lilium). Converge's evidence-first design is that
+  discipline applied to model weights.
+- Built the full stack, from the merge protocol and its CRDT substrate to the evaluation and evidence system.
+- Ran the research programme behind every published measurement.
+- Author of the paper on CRDT-based model merging ([arXiv:2605.19373](https://arxiv.org/abs/2605.19373)).
+
+<sub>Current releases are licensed for non-commercial research, a condition inherited from an upstream source model.
+Each model card states its licence in full.</sub>

index.html

@@ -0,0 +1,67 @@
+<!doctype html>
+<html lang="en">
+<head>
+<meta charset="utf-8">
+<meta name="viewport" content="width=device-width, initial-scale=1">
+<title>Optitransfer · converge</title>
+<meta name="description" content="Converge: building frontier-level open models from the work the open-source community has already published. One repeatable process at every model class, with every result published with per-item evidence.">
+<style>
+  :root { --bg: #ffffff; --fg: #1a1d24; --muted: #5b6270; --line: #e6e8ec; --accent: #3b4fd8; }
+  @media (prefers-color-scheme: dark) { :root { --bg: #0f1115; --fg: #e8eaf0; --muted: #9aa1ae; --line: #252a33; --accent: #8ea0ff; } }
+  * { box-sizing: border-box; }
+  body { margin: 0; background: var(--bg); color: var(--fg); font: 16px/1.6 system-ui, -apple-system, "Segoe UI", sans-serif; }
+  main { max-width: 760px; margin: 0 auto; padding: 56px 16px 72px; }
+  h1 { font-size: 2.2rem; margin: 0 0 .25rem; letter-spacing: -.01em; }
+  .lede { font-size: 1.15rem; margin: 0 0 .25rem; }
+  .sub { color: var(--muted); margin: 0 0 2rem; }
+  h2 { font-size: 1.05rem; text-transform: uppercase; letter-spacing: .06em; color: var(--muted); margin: 2.25rem 0 .5rem; }
+  a { color: var(--accent); text-decoration: none; } a:hover { text-decoration: underline; }
+  ul, ol { padding-left: 1.2rem; } li { margin: .3rem 0; }
+  .rel { border-top: 1px solid var(--line); }
+  .rel div { display: flex; gap: 1rem; padding: .6rem 0; border-bottom: 1px solid var(--line); }
+  .rel b { min-width: 9.5rem; flex-shrink: 0; }
+  footer { margin-top: 3rem; color: var(--muted); font-size: .9rem; }
+  @media (max-width: 520px) { .rel div { flex-direction: column; gap: .1rem; } }
+</style>
+</head>
+<body>
+<main>
+  <h1>Optitransfer</h1>
+  <p class="lede"><strong>Converge:</strong> building frontier-level open models from the work the open-source community has already published.</p>
+  <p class="sub">One repeatable process at every model class. Every result is measured against its base model and published with per-item evidence.</p>
+
+  <h2>Purpose</h2>
+  <p>The community has published thousands of fine-tuned models, each adding some skill to a shared base. Converge brings that work together into a single open model for each class, without pre-training a new one. It measures every change, keeps every gain and publishes every regression.</p>
+
+  <h2>The goal</h2>
+  <p>Reach frontier-level performance in every model class: 7B first, then each larger size class and other model families. The model must improve on every axis without a material regression against its base, until no public fine-tune improves any axis further.</p>
+
+  <h2>The process</h2>
+  <ol>
+    <li><strong>Start</strong> from a strong open base model in the class.</li>
+    <li><strong>Discover</strong> the compatible fine-tunes the community has published.</li>
+    <li><strong>Assimilate</strong> them into one model.</li>
+    <li><strong>Measure</strong> every axis against the base on identical items, on public protocols.</li>
+    <li><strong>Release</strong> a model only when it holds on the whole board, and publish every regression. This is the release rule from the corrected releases on.</li>
+    <li><strong>Repeat</strong>, then carry the process to the next class.</li>
+  </ol>
+  <p>The construction method is proprietary. The evaluation is fully public.</p>
+
+  <h2>Releases</h2>
+  <div class="rel">
+    <div><b><a href="https://huggingface.co/Optitransfer/Qwen2.5-7B-Instruct-converge-collective-v1">converge v1</a></b><span>Public. The first 7B release. It carries a corrected merge defect; see its card.</span></div>
+    <div><b><a href="https://huggingface.co/Optitransfer/Qwen2.5-7B-Instruct-converge-collective-v2">converge v2</a></b><span>Public. The current release. It is built on v1, so it inherits the defect.</span></div>
+    <div><b>corrected v1 and v2</b><span>Same recipes, defect removed. Being benchmarked across the full board now.</span></div>
+    <div><b>converge v3</b><span>Next. Released only once it holds against its base and the previous release.</span></div>
+  </div>
+
+  <h2>Open components</h2>
+  <ul>
+    <li><a href="https://github.com/mgillr/crdt-merge">crdt-merge</a>: conflict-free merging for data, models and agents, with the E4 trust engine.</li>
+    <li><a href="https://github.com/mgillr/acfa-rs">acfa-rs</a>: accountable federated aggregation with offline-verifiable receipts.</li>
+  </ul>
+
+  <footer>Ryan Gillespie, founder · <a href="https://huggingface.co/Optitransfer">@Optitransfer</a></footer>
+</main>
+</body>
+</html>