Skip to content

GenoLeWM checkpoint and evidence bundle

Space Checkpoint Dataset GitHub

Checkpoint validity notice (2026-07-10): Every checkpoint documented on this page, including the stable v0.1 package and the v0.2.1 run-tree model, uses the legacy_raw_v1 state contract. Although the configs declared encoder.normalize: true, the released implementation trained and evaluated on raw pooled Carbon states. Training sources used global pooling while targets were edit-centered, but every historical centered pool was shifted one hidden token left because the leading <dna> control token was omitted from the coordinate conversion; rollout/candidate centers differed and cache v1 omitted that identity. The pinned Carbon tokenizer.py also performed an unpinned, network-capable Qwen/Qwen3-4B-Base lookup, so its hash did not establish a self-contained runtime identity. The v0.2.1 Phase 2 KL was computed from frozen target states; it changed the reported scalar loss but was constant with respect to the trainable predictor and action encoder, so it supplied no gradient. The metrics below are retained only as historical implementation outputs. These defects invalidate the checkpoints and metrics as evidence for the intended normalized method. Corrected results require fresh training and evaluation in a new l2_normalized_v2 checkpoint lineage. The repaired source loader uses local pure-DNA tokenization and validated token layout, but no checkpoint has yet supplied model-quality evidence under those repairs.

GenoLeWM is an alpha research system for action-conditioned latent world models over genomic edits. This model repository publishes the stable v0.1 checkpoint package: trainable GenoLeWM predictor and action-encoder weights, calibration data, manifest-bound hashes, training evidence, evaluation evidence, and release checksums.

This is not a standard transformers.AutoModel.from_pretrained() checkpoint. Load it through the geno-lewm runtime. Carbon-500M is a frozen state encoder dependency and is resolved separately from this repository.

Claim Boundary

Use this checkpoint for historical replay, local artifact inspection, and method development. Do not use it for scientific scoring or as a diagnostic model, clinical decision-support system, deployment-readiness claim, privacy claim, or evidence of broad superiority over Carbon. The measured results below are historical evaluations of the legacy_raw_v1 implementation. They are not valid evidence about the intended normalized GenoLeWM method.

No clinical utility claim is made.

At A Glance

Item Value
Stable package in this repo geno-lewm-v0.1.0-r1
Newer run-tree checkpoint used by the Space geno-lewm-v0.2.1-r1 under abdelstark/geno-lewm-runs
State contract for both published checkpoints legacy_raw_v1
Runtime Historical geno-lewm + Carbon path; not self-contained because its pinned tokenizer made an unpinned transitive Qwen lookup
Historical task Single-SNV residual output over reference windows
Scientific status Historical implementation outputs; normalized-method results not yet published
Strongest honest conclusion Reproducible legacy systems artifact, not a valid evaluation of the intended normalized method
Not supported Clinical utility, deployment readiness, privacy assurance, runtime assurance, broad model superiority

Which Checkpoint Should I Use?

Target Location Use when
Stable package abdelstark/geno-lewm You need checksum-bound replay of the historical v0.1 legacy_raw_v1 artifact. Do not use it to evaluate the normalized method.
Space/default demo checkpoint geno-lewm-v021-strong-4f36eef-10k-r1/suite/model You need historical v0.2.1 artifact replay or UI compatibility. Its benchmark and planning outputs do not evaluate the normalized method.
Artifact console abdelstark/geno-lewm Space You want to inspect historical artifacts and checkpoint metadata. Legacy scientific scoring is disabled.

Published Artifacts

Artifact Location Notes
Stable model package this repository Public package geno-lewm-v0.1.0-r1.
Generated package card model_card.md Checksum-bound output from tools.release.model_package; kept terse so it can be re-rendered exactly.
Manifest manifest.json Canonical identity for weights, calibration, config, encoder, and eval report.
Package checksums SHA256SUMS Hashes for the v0.1 release package files.
Training evidence training_run_manifest.json, training_run_card.md, training_run_SHA256SUMS Carbon-backed training-run metadata and hashes.
Evaluation evidence eval_metrics.json, eval_report.md, eval_config.effective.yaml Held-out chr21 ClinVar evaluation for v0.1.
Efficiency evidence efficiency_report.json Release efficiency measurement for v0.1.
Dataset package abdelstark/geno-lewm-data Public data snapshot and data card.
v0.2.1 run tree abdelstark/geno-lewm-runs Newer checkpoint, broader benchmark suite, planning demo, and generated paper.
Historical generated paper paper.serious-completion.md Superseded manuscript containing the withdrawn negative-result interpretation; retained only for provenance.

Model Identity

Field Value
Release id geno-lewm-v0.1.0-r1
Model version 0.1.0
Manifest id sha256:861ec142cc87f3fac01751ef538553356dfba439e6da99064b4adb121e75c215
Predictor artifact predictor.safetensors
Predictor hash sha256:6642c604a1352727969c86664f291fd6d2193c1c65bc6f9baf9b716469c52731
Action encoder hash sha256:8b2311d768855ab440b26dbbef5ddbda252cc8bb2c69509d28fa4bcf8eff025a
Calibration hash sha256:d4cf4778ac8e5557d363aca43cd13723b0ed9983b83215ab164d2b642b886201
Frozen encoder Carbon-500M, mounted as /carbon in release jobs
Encoder revision 5d31d59b3c845b288a13aedb1358934196852eec
Effective state contract legacy_raw_v1
Dataset snapshot geno-lewm-data-v0.1.0-r1

The newer Space default checkpoint has model id sha256:cddb8f3b9671090201370b9824b9da741b933ff296b651238f022df5f3ed6af4 and lives in the v0.2.1 run tree. It also uses legacy_raw_v1; it is historical run evidence, not a corrected replacement for the stable v0.1 package.

Method

GenoLeWM treats an edit as an action over a frozen genomic state embedding. For a reference window, Carbon-500M encodes the pre-edit state. GenoLeWM's trainable action encoder represents the edit, and the predictor estimates the post-edit latent state. Under a valid same-contract lineage, candidate surprise scores can be derived from the relationship between predicted and observed post-edit states and calibrated with a matching artifact. The published legacy residuals and calibration do not satisfy that contract.

The package intentionally separates:

  • frozen state encoder: Carbon-500M, loaded separately;
  • trainable GenoLeWM weights: predictor.safetensors and action_encoder.safetensors;
  • release identity: manifest.json and SHA256SUMS;
  • evidence: training, evaluation, efficiency, and run-tree reports.

The intended method applies L2 normalization to Carbon state vectors before training, scoring, rollout evaluation, and planning. Published checkpoints did not apply that transformation. Compatibility therefore preserves them as legacy_raw_v1; corrected runs must declare l2_normalized_v2 and use a new checkpoint identity rather than reinterpret old weights or metrics.

The same boundary applies to tokenization and coordinates. The released custom Carbon tokenizer could resolve Qwen/Qwen3-4B-Base over the network without a pinned revision, and historical centered pooling ignored the leading <dna> hidden token. Current source repairs both paths with a local pure-DNA tokenizer and token-ID-derived centers. They are implementation invariants, not evidence that a corrected model has learned a valid transition.

Install And Load

Install the package:

python -m pip install "geno-lewm[train,eval]==0.2.1"

Download the stable v0.1 model package:

from huggingface_hub import snapshot_download

model_dir = snapshot_download("abdelstark/geno-lewm")

Download the newer v0.2.1 run-tree model used by the Space:

from pathlib import Path
from huggingface_hub import snapshot_download

prefix = "geno-lewm-v021-strong-4f36eef-10k-r1/suite/model"
run_root = snapshot_download(
    "abdelstark/geno-lewm-runs",
    allow_patterns=f"{prefix}/*",
)
model_dir = Path(run_root) / prefix

Carbon-500M must also be available. The release manifests record the encoder as /carbon because training, evaluation, and demo jobs mounted HuggingFaceBio/Carbon-500M there at revision 5d31d59b3c845b288a13aedb1358934196852eec. The Space no longer resolves this encoder for scoring; its published-checkpoint surface is inspection-only.

Scoring Contract

The runtime preserves this input-validation contract for compatibility, but running a published legacy_raw_v1 checkpoint reproduces invalid historical residual output rather than a scientific score. The hosted Space therefore disables scoring. The REF allele in chrom:pos:ref:alt must match the supplied reference window at the one-based locus implied by pos and --window-start-bp. If it does not match, scoring raises WindowMismatchError before model inference.

Historical replay example with a synthetic reference-matched window:

geno-lewm-score \
  --model-dir "$MODEL_DIR" \
  --backend auto \
  --variant chrSynthetic:3073:A:T \
  --window "$(python - <<'PY'
print("ACGT" * 3072)
PY
)" \
  --window-start-bp 0 \
  --receipt receipt.json

Do not interpret the output as edit effect, surprise, pathogenicity, or model quality. Corrected scientific scoring requires a newly trained l2_normalized_v2 checkpoint and matching calibration table.

Historical v0.1 Training Summary

The stable v0.1 checkpoint was trained as a JEPA-style predictor over frozen Carbon-500M raw pooled latent states under legacy_raw_v1.

Field Value
Run id first-snv-carbon-500m-r1
Config training_config.effective.yaml
Commit cd2bfccb33ec5a2df3c4707e8be8443f4682dad3
Samples 160,000
Steps 20,000
Final training loss 0.36124
Status completed

Historical v0.1 Evaluation

Held-out ClinVar GRCh38 chr21, binary P/LP versus B/LB labels. Scores use sigma_raw; intervals are deterministic stratified bootstrap confidence intervals from eval_metrics.json. These values measure the released legacy implementation; they do not evaluate the intended normalized method.

Split N Positives Negatives Metric Value 95% CI
eval_clinvar_chr21 3,000 494 2,506 AUROC 0.519160 0.491366 to 0.546846
eval_clinvar_chr21 3,000 494 2,506 Average precision 0.165174 0.155331 to 0.177035
eval_clinvar_chr21 3,000 494 2,506 Balanced accuracy at 0.5 0.500000 0.500000 to 0.500000
eval_clinvar_chr21 3,000 494 2,506 Accuracy at 0.5 0.164667 0.164667 to 0.164667

Interpretation: this v0.1 slice does not establish useful clinical performance, non-coding performance, multi-edit behavior, or superiority over Carbon.

v0.1 Efficiency

Measured by tools.release.efficiency_report on cuda:NVIDIA H200.

Measurement Value
Single-variant latency 494.056 ms
Batched throughput 2.024 variants/s
Peak memory 1,152,656,384 bytes

v0.2.1 Benchmark Evidence

The table is retained as historical legacy_raw_v1 output. The v0.2.1 run tree includes broader benchmark coverage and Carbon zero-shot comparisons, but the results do not evaluate the intended normalized method. In addition, its Phase 2 KL was constant with respect to every optimized parameter and supplied no gradient, so the claimed active regularization was absent from optimization.

Slice N Metric GenoLeWM Baseline Delta
ClinVar coding 16 AUROC 0.734375 0.921875 -0.187500
ClinVar coding 16 Average precision 0.852976 0.951923 -0.098947
ClinVar coding 16 Balanced accuracy 0.750000 0.687500 +0.062500
ClinVar non-coding 16 AUROC 0.562500 0.875000 -0.312500
ClinVar non-coding 16 Average precision 0.605456 0.914423 -0.308967
ClinVar non-coding 16 Balanced accuracy 0.437500 0.687500 -0.250000
BRCA2 saturation 32 Spearman rho 0.149194 0.476906 -0.327713
TraitGym Mendelian 32 Spearman rho -0.027965 -0.083894 +0.055929
Phased-haplotype rollout 8 Cosine mean 0.288861 0.997831 -0.708970
Synthetic edit-chain rollout 8 Cosine mean 0.301608 0.991240 -0.689631

The v0.2.1 readiness report is ok=true for artifact coverage and provenance, not for scientific validity of the normalized method. The autoregressive rollout speed report is ok=false: K=5 measured 2.41x speedup against a 2x target, while K=20 measured 2.47x against a 5x target and missed the target.

The v0.2.1 efficiency report measured one sample with no warmup on cuda:NVIDIA H200: 115,262.94 ms single-variant latency, 0.3095 variants/s throughput, and 1,966,149,632 bytes peak memory. Treat that as run evidence, not a production serving benchmark.

Planning Demo Evidence

The v0.2.1 run tree includes a deterministic synthetic multi-SNV planning demo. It ran with manifest-backed runtime evaluation and recorded 384 evaluations, best_distance=23.656930390534644, and stopped_reason=patience.

This is historical legacy_raw_v1 execution evidence, not useful-planning or normalized-method evidence. It shows only that the planner, legacy checkpoint, window artifacts, and receipt path ran end to end.

Troubleshooting

Symptom Likely cause Fix
WindowMismatchError: window bases do not match edit.ref_bases at locus The variant REF allele does not match the supplied window at the requested coordinate. Use a FASTA-backed VCF path or correct --window, --window-start-bp, and the variant REF allele so they describe the same reference sequence.
Space has no scoring controls Legacy scientific scoring is intentionally disabled after the state-contract audit. Use the Space for artifact inspection; wait for a fresh l2_normalized_v2 checkpoint before scientific scoring.
AutoModel.from_pretrained("abdelstark/geno-lewm") fails This repo is a GenoLeWM package, not a native Transformers architecture repo. Use snapshot_download plus geno-lewm-score or the geno_lewm.deploy runtime.

Limitations

  • Alpha research checkpoint; not a clinical, diagnostic, or deployment model.
  • All published checkpoints use legacy_raw_v1; no corrected l2_normalized_v2 result is published.
  • The v0.2.1 Phase 2 KL supplied no gradient to trainable parameters.
  • v0.1 evaluation is narrow: held-out chr21 ClinVar P/LP versus B/LB labels.
  • v0.1 and v0.2.1 metrics are retained only as historical implementation outputs and must not be cited as evaluation of the intended normalized method.
  • Carbon-500M is required at runtime and is resolved separately from this model package.
  • Calibration is proof-scale and should be interpreted only within the reported artifact context.
  • Fixture outputs and UI demos are not model-quality evidence.

Citation And Reports

License

Apache-2.0 for GenoLeWM source and metadata. Upstream Carbon-500M and dataset terms apply to their respective artifacts.