LatticeLM public research archive
The canonical LatticeLM checkpoint has 48,636,168 parameters and was trained from random initialization for 1,900,000,000 tokens. This curated repository preserves the exact original training checkpoint, an exactly equivalent safetensors inference export, tokenizer/config, evaluations, provenance, and selected development checkpoints.
Canonical model
Co4 causal architecture: 12 layers, width 552, FFN width 1728, 6 query heads / 2 KV heads, QK normalization, untied embeddings, vocabulary 4096, context 256. The optimizer is Muon hybrid with a fresh WSD schedule. DATA-D-v4 combines FineWeb-Edu, Wikipedia, FineWeb and FineMath; source revisions, composition, hashes, filtering and contamination metadata are preserved in manifests. Natural corpus content is withheld pending redistribution review.
The original final report records held-out loss 2.63049, WikiText perplexity 21.9298 and BPB 1.41346, ARC-Easy 0.36785, HellaSwag 0.27096, PIQA 0.56692 and WinoGrande 0.50987. See the production reports for evaluation protocols and limits. This small model is intended for research; its short context and limited training scale constrain generation quality. These results do not establish reliability for high-stakes use.
| Artifact | SHA-256 |
|---|---|
| Native final checkpoint | 94b91a2d39c2889209cb47b2e04b2c0d307a25610c40c5a124b595020eb762b0 |
| Tokenizer | 294925d61198dfa0863f91193349f6ebead420d06a3fb650a5122d2264a4ebf7 |
| DATA-D-v4 manifest | 9d2811cfb2824b0dd541af89ebfb64403b3ec7143db97d52c4046aafb8eeae7f |
The canonical configuration in config/final_model_config.json is extracted from the exact 48.6M checkpoint and restore-tested against it. A historically named local config described an earlier 8M model; that original remains in the broad archive as historical evidence.
Development artifacts
Development checkpoints may be incomplete, negative, exploratory or superseded. They are preserved for transparency and continued research and are not all recommended for deployment. Selected R32/R16 controls, Differential-Co4/nCo4 artifacts, recovered 8M control provenance and campaign reports are included. The four historical 8M controls are retained with recovered exact trainer snapshots and independent certificates.
The fresh 32M dense run completed 300M tokens on October 11, 2026. The active run at evacuation was instead F8-CO4-FULLCONTROL-50M-S1337; with user confirmation it was truncated at 4,268,032 / 50,000,000 tokens because GCP credits ran low. Its exact model/optimizer/RNG/sampler state passed strict resume validation and its checkpoint was restored and hash-verified. Its receipt records the stable WSD region and LR 0.0008. It is not a completed 50M result.
Loading and reproducibility
Use the source branch at 1293deb76d199050482182f5f2ac0f52f9548353 and the environment version receipt in reports/evacuation/. The exact original production source is also preserved under source-snapshots/canonical-6e2c6f0f6f5278a2991c04708859baae7580368c/.
import torch
from huggingface_hub import hf_hub_download
from tokenizers import Tokenizer
from latticelm.config import LatticeConfig
from latticelm.model import build_model
repo = "insightlabs38-pixel/LatticeLM-public-archive"
path = hf_hub_download(repo, "canonical/checkpoint/final.pt")
checkpoint = torch.load(path, map_location="cpu", weights_only=False)
model = build_model(LatticeConfig(**checkpoint["config"]))
model.load_state_dict(checkpoint["model"], strict=True)
model.eval()
tokenizer = Tokenizer.from_file(hf_hub_download(repo, "tokenizer/final_corpus_4k.json"))
tokens = torch.tensor([tokenizer.encode("LatticeLM research.").ids])
with torch.no_grad():
logits, _ = model(tokens)
The public VM archive contains the broader research history, native checkpoints, controller state, logs, exact source/config snapshots, synthetic datasets, and detailed inventories. Public bucket URI: hf://buckets/insightlabs38-pixel/latticelm-vm-archive. Archive date: 2026-10-12T21:11:37.997469+00:00. The old private research repository remains private.
License status
No general license has been invented for the canonical model. Author review is still required to select its license. IHMS-derived development implementation preserves its existing CC-BY-NC-4.0 attribution and complete upstream license in NOTICE / LICENSES/. This does not apply a blanket license to unrelated artifacts. Natural corpora and benchmark examples are represented by metadata and reconstruction scripts until redistribution obligations are resolved.