v0.3.0-v4-preview - V4 architecture (untrained preview)

V4 architecture preview — untrained.

This pre-release publishes the V4 graph-native architecture (Unified Dynamic Token-Concept-Relation Graph Causal LM) as code + design docs only. There is no trained checkpoint: V4 is component-smoke-tested but not validated as a working language model. The trained, released line remains Stage C / V3 (see v0.2.0-stagec).

Included:

• V4 model / dataset / training entry + 8 contract smoke tests
• V4 teacher-labeling + audit pipeline
• 20 V4 design documents + architecture paper
• Browser-based training-data audit UI

Shared for architecture review and collaboration.

v0.2.0 Stage C checkpoint package

TMCRA TokenGraph-LLM Stage C / Dynamic Token Graph Decoder V3

This release provides the Stage C public model package for TMCRA TokenGraph-LLM. Release assets are model artifacts only. Algorithm source, graph builders, full-chain training scripts, and evaluation templates are maintained in the GitHub source repository on main.

Source repository: https://github.com/reshuibuduo/TMCRA-TokenGraph-LLM
Hugging Face model page: https://huggingface.co/2009YU/TMCRA-TokenGraph-LLM

Model

• Asset: tmcra_tokengraph_stagec_model_package_20260606.zip
• SHA256: cc23285628eaed47c20009b6be6b5eb0600ded57ac2e09519370d97158fecd33
• Checkpoint: token_graph_dynamic_decoder_v3.pt
• Parameters: about 114.6M
• Architecture: dim=512, graph_layers=8, decoder_layers=10, untied embeddings
• Training: bf16, about 1.03M effective token-graph samples, 62,000 steps

Package contents

• checkpoint/token_graph_dynamic_decoder_v3.pt
• dataset_metadata/tokenizer.json
• dataset_metadata/manifest.json
• training_summary_stagec_public.json
• docs/ technical overview and smoke benchmark notes
• SHA256SUMS.txt

Full-chain source update

The source repository now includes:

• open-corpus schema2 conversion scripts;
• optional semantic teacher annotation scripts;
• token graph dataset builders;
• Stage C full-chain and sharded training templates;
• graph ablation and token attribution evaluation.

Start with:

• docs/FULL_CHAIN_TRAINING.md
• docs/FULL_CHAIN_TRAINING_ZH.md
• scripts/run_stagec_full_chain_template.sh
• scripts/run_stagec_sharded_training_template.sh

Smoke results

• Stage C normal: total loss 6.512117, lm loss 4.641285
• Stage C no_edges: total loss 8.310654, lm loss 5.790666
• Stage C shuffle_edges: total loss 7.702783, lm loss 5.169387
• TinyStories smoke: normal avg words 73.88
• BLiMP smoke: 59%-64% on three small likelihood subsets

These are smoke tests, not leaderboard claims. Stage C is still an early graph-native language model prototype, not a production LLM.

The previous v0.1.0-prototype release is now considered the legacy small prototype checkpoint.

v0.1.0 prototype checkpoint package

Prototype checkpoint package for TMCRA TokenGraph-LLM.

Includes:

• token_graph_llm_v1_relation_finetune_3k.pt
• tokenizer.json
• model card, bilingual README, manifest, SHA256 file

This is an early research prototype checkpoint, not a polished production model.