GoldWorm v6 β€” Byte-Level Causal Transformer (2.76M params)

Verified result: full BPC 1.007349 (step 8000, TinyStories-valid, 129,429 fixed outer-val positions). Beats GPT-2 124M BYTE_BPC = 1.142985 (TOKEN_BPC = 9.152823) on the same metric.

Dataset: roneneldan/TinyStories (TinyStories-valid.txt, 19.4 MB, sha256 94e431816c4cce81...). Protocol: windowed stride-16 ctx-256 over fixed evaluation set; determinism verified (two bit-identical passes, max_abs_diff_per_position_bits = 0.0).

Model facts

  • Type: ByteTransformer
  • Vocab: 256 (byte-level)
  • Config: vocab_size=256, d_model=192, n_layer=6, n_head=4, ffn_mult=4, ctx=256, tie_embeddings=True, dropout=0.0
  • Params: 2,757,504
  • Checkpoint: run/andrew_8000/checkpoint_selected.pt β†’ model.safetensors
  • Selection (inner-val): step 8000, inner_val_windowed_bpc = 1.051187
  • Full-val (outer): step 8000, full_bpc = 1.007349
  • Baselines: unigram 4.437 / bigram 3.317 (add-k 0.01, train-matched)
  • Gate: well below frozen v1 reference (2.566)

Files

  • model.safetensors β€” exported weights (tied embeddings cloned; load with ByteTransformer(config))
  • config.json β€” architecture config
  • eval_results.json β€” full outer-val result
  • demo/app.py β€” Streamlit text-generation demo
  • demo/requirements.txt β€” torch, safetensors, streamlit, numpy

Use with Streamlit demo

pip install -r demo/requirements.txt
streamlit run demo/app.py

Prompt is drawn from corpus/corpus.txt outer-val region; generation uses temperature / top-k sampling.

How to load (Python)

import sys, torch
sys.path.insert(0, '.')
from gwlm_v6 import TransformerConfig, ByteTransformer
from safetensors.torch import load_file

cfg = TransformerConfig(vocab_size=256, d_model=192, n_layer=6, n_head=4, ffn_mult=4, ctx=256, tie_embeddings=True, dropout=0.0)
model = ByteTransformer(cfg)
state = load_file("model.safetensors")
model.load_state_dict(state)
model.eval()

Verification

  • Chunk shape / parity / replication checks verified per LLM_PARITY_ROADMAP.md
  • Independent replication (step 8000) confirmed bit-exact with selection frozen before any outer_val read
  • Source repo: C:/Users/Student/GoldWorm/GoldWorm/goldworm-lm-v0/
Downloads last month
231
Safetensors
Model size
2.81M params
Tensor type
F32
Β·
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support