REAX / COMPETITIONS / GENERALIST / LINEAGE
Lineage check
See how intake checks your model’s weight sources and how to run the public check. This page backs section 3 of the round 2 rulebook (checker version 1.6.2).
At a glance
- Checker version. 1.6.2.
- Script hash (
check_script_sha256).302932efe9a4d4f77a358e9bdcc098fff87665f02557e021fcac0858694cf3f1. It is not the hash of one file; see Check the script. - Script. github.com/reaxlabs/reax-subnet, tools/reax-lineage.
- Record schema. reax-lineage.schema.json, the format of the
reax-lineage.jsonfile in your repo (rule 3.1). - Engine templates. ENGINE-TEMPLATES.md, the architectures and readouts the open engine serves (rule 2.4).
- Rules. Sections 3.1 to 3.9 of the round 2 rulebook; a lineage result is final at the objection deadline of the first window in which the entry is scored (3.9).
Method
The check runs on your committed revision before any scoring. It aims to finish within about 6 hours of your commitment; that is a target, not a deadline (rule 3.2). It does two kinds of work.
- Your record and files. The on-chain commitment pins the model and the record. The record gives your committing hotkey and pinned weight parents, or
from_scratch. The checker ignoresentry.revisionin older records. You can omit it. The check verifies your parents and that your weights and config can be served by a supported engine template. Repo model IDs can be evidence for weight comparisons. - Your weights. It does not download the whole model. It reads the safetensors headers and a small sample of rows by HTTP range requests (next section). It takes your change, your weights minus your declared parent or parents, and compares it with the change of every earlier model of the same architecture: the list, the public releases of that family, every earlier entry of this round (all revisions, also withdrawn and still private ones), round 1's entries, REAX releases, the top rows and all flagged models of that family on the Hugging Face Decision Index, its most-downloaded public fine-tunes, and every model named in your repo.
- Two numbers per candidate. For each sampled tensor it computes beta (how much of the other model's change is contained in yours) and cos (how similar the two changes are in direction). It takes the median per layer, then summarises across layers: the median beta and cos, the share of layers with a beta of at least 0.025, and a z value (beta divided by its standard error across layers). A null level, measured with the layers shifted against each other, guards against chance matches.
- Declared parents must be found. With one parent, the check confirms that your weights sit in the frame of that parent. With several, it fits your change jointly on the parents' changes; every declared parent needs a fitted weight whose lower bound is above zero on every sample.
- Other weight sources. A model of the same base family that the check finds in your weights but your record does not name is a weight overlap. It is never a penalty and needs no correction: at most the three overlaps with the largest positive weight in a joint fit of your own change join your bar list (rule 3.6). Honest fine-tunes of one base on the same public data look alike in their weights, and the check cannot tell such siblings from a merge.
- Disguises. If the frame check fails, a second test (Tier B) tries to undo a rotation, permutation or rescaling of your parent, which reads four whole layers and 16,384 embedding rows.
- Public and sealed sample. The check uses two samples, A and B. Sample A is published; Sample B uses a sealed seed that we do not publish. A dry run uses Sample A only, so a dry-run PASS is not a promise (rule 3.2).
Tensor sample
A fingerprint is the weight sample of one model. What is sampled, in reax_lineage/fp.py:
- Layer weights. Every 2-D floating-point tensor with at least 1,000,000 elements inside the repeated language-model layer blocks, keyed by layer number and tensor name. From each, 4 consecutive rows at a position fixed by the sha256 of the seed, the tensor name and the layer.
- Mixture-of-experts models. The router of every layer in full, and 4 experts per layer at seeded indices.
- Output head and embeddings. 1,024 token rows of each, taken as 32 blocks of 32 rows from the first 100,000 rows.
- Not sampled. The vision tower, norms and biases.
- Public seed. Sample A uses the seed
A0(sha256aa508c2187fca56f397ff75adc52b94e02f38122cdd48bd42105106e5e0f8e14). This is the sample you can reproduce. Sample B's seed is sealed; results carry only its sha256.
Thresholds
All numbers are in reax_lineage/thresholds.json, which is part of the script hash. A model is found by the whole-model test when beta, cos, share of layers, z and the null condition all hold, or by the splice test when at least splice_layers layers match closely.
| Name | Value | What it means |
|---|---|---|
beta_min | 0.05 | Whole-model test: the other model's change is in yours with a coefficient (beta) of at least this. |
cos_min | 0.01 | Whole-model test: the cosine between the two changes is at least this. |
share_min | 0.80 | Whole-model test: at least this share of layers have a beta of at least share_layer_beta. |
share_layer_beta | 0.025 | The beta a layer needs to count towards share_min. |
z_min | 15.0 | Whole-model test: beta divided by its standard error across layers is at least this. |
null_x | 5.0 | Whole-model test: beta is at least this many times the null level (the largest |beta| found when the layers are shifted by null_shifts). |
null_shifts | 1, 2, 3, 5, 8, 13, 21 | The layer shifts used for the null level. |
splice_layers | 2 | Splice test: a model is also found if at least this many layers match it closely. |
splice_beta | 0.5 | Splice test: a layer matches closely with a beta of at least this … |
splice_cos | 0.5 | … and a cosine of at least this … |
splice_min_tensors | 4 | … over at least this many tensors of the layer. |
frame_cos | 0.80 | Declared parent confirmed: a sampled row counts when its cosine with the parent is at least this. |
frame_share | 0.97 | Declared parent confirmed: at least this share of rows count. |
scale_tol_log2 | 0.05 | Declared parent confirmed: a row is rescaled when |log2 of its scale| is above this. |
scale_share_max | 0.05 | Declared parent confirmed: at most this share of rows may be rescaled. |
start_beta | 0.5 | Declared parent confirmed: the entry's change measured against the parent shows a beta of at least this. |
tierb_cos | 0.80 | Rotation / permutation test (Tier B): the cosine a row needs after the transform is undone. |
copy_share | 0.999 | Copy (rule 3.5): at least this share of sampled layer values equal the other model's within bf16 rounding. |
head_equal | 0.999 | Copy: the same for the sampled output-head and embedding rows. |
bf16_tol | 0.00390625 (2-8) | The relative tolerance of "equal within bf16 rounding". |
grid_share | 0.90 | Quantisation round trip: at least this share of 128-value groups lie on a quantisation grid while the parent's do not. |
bar_list_max_overlaps | 3 | At most this many weight overlaps join the bar list (rule 3.6). |
deliberate_share | 0.5 | Shown on the uid page for reference only; it decides nothing. |
beta_action | 0.25 | Documented value; no verdict uses it. |
What the check can say
- PASS. Published with your bar list; scoring may start (rule 3.3).
- NEEDS CORRECTION. A problem blocks scoring, such as an unverifiable parent declaration or a model that cannot be scored with the supported template. Other issues can be reported as notes. No penalty. You have 24 hours to send a corrected record; if you think the check is wrong, dispute it.
- REFUSED. Something not allowed: for example a parent that is not allowed, more than 4 parents or a share below 0.10, parents served by different templates, a copy, or a quantisation round trip that is not declared as quantisation-aware training. No penalty; you may commit a new revision.
- Misdeclaration candidate. Evidence suggests that a miner deliberately hid a parent or disguised another holder’s weights. The script only prepares the finding with its measurements; a person decides (rule 3.8).
A technical failure on our side or at Hugging Face never refuses or corrects an entry. A repo we cannot read is retried hourly (rule 2.7). Result codes and their machine-readable fields are documented in the README of the script.
Check the script
check_script_sha256 is printed in every result file. It is computed by script_sha256() in reax_lineage/check.py as one sha256 over: for each file of reax_lineage/ whose name ends in .py or is thresholds.json, in sorted order, the file name and then the file bytes; then the bytes of schema/reax-lineage.schema.json and schema/reax-lineage-1.schema.json. It does not cover bin/, tests/ or calibration/. To check a copy of the script:
cd tools/reax-lineage
bin/reax-lineage --version # reax-lineage 1.6.2
python3 -c "from reax_lineage.check import script_sha256; print(script_sha256())"
# 302932efe9a4d4f77a358e9bdcc098fff87665f02557e021fcac0858694cf3f1Run it yourself
You can run the check on the public sample before you commit. When the signed MCP dry run opens, it will check public repos (3 per hotkey per day). Check competition status for availability. For a private repo, or to see the numbers, run the published script on your own weights. It needs Python 3 with a pip-wheel numpy, and a read-only Hugging Face token only if you read repos from the Hub.
git clone https://github.com/reaxlabs/reax-subnet
git -C reax-subnet checkout 1d6298f9295502bb84a3fb907a09531551f6339a
cd reax-subnet/tools/reax-lineage
python3 -m venv venv && venv/bin/pip install -r requirements.txt
export REAX_LINEAGE_PYTHON=$PWD/venv/bin/python
# the Sample A fingerprint of a local checkpoint (seed A0)
bin/reax-lineage fingerprint --local /path/to/checkpoint --name my/model --cache ./cache --out my-model.npz
# the dry run: Sample A only, public models and revealed entries only
bin/reax-lineage dry-run --local /path/to/checkpoint --repo my/model --rev REVISION \
--hotkey YOUR_HOTKEY --block COMMIT_BLOCK --record reax-lineage.json \
--list list.json --comparison-db ./db --cache ./cache --out result.json
The script needs two inputs: list.json, the list at your commitment time, and a comparison database (index.json plus fingerprint files) of the models to compare with; calibration/build_db.py shows how one is built from fingerprints. The result is result.json. A run that ends without a result prints one JSON line with an error_kind and exits with code 2; the README of the script explains each kind. The intake check also uses the sealed sample, so a PASS here is not a promise.
Rules: Competition rules · Miner FAQ · Competitions · Record schema · Engine templates · Script