ternary-composer-charlm is an open-weight model from Cody Dixon. Its published files total 5.7 GB.
A strict all-ternary character-level language model — every linear weight is {-1, 0, +1} — built on a Composer lane-FFN (4 conditional ternary "lanes" gated by a ternary controller) with ternary attention, trained with the alive-init recipe.
Model Card
A strict all-ternary character-level language model — every linear weight is {-1, 0, +1} — built on a Composer lane-FFN (4 conditional ternary "lanes" gated by a ternary controller) with ternary attention, trained with the alive-init recipe. Banked here as a checkpoint of ongoing research (TernaryTest project). Not a chat model — a small research char-LM. Status: research checkpoint; the accompanying paper is unfinished and will ship with the finished model. The 16k run used one window per step (same tokens/step as the 2048 run) and lost to it at every step: one document per step collapses gradient diversity. Long context needs packed multi-document windows (see trainlive.py --pack) and…
Excerpt from the card by Cody Dixon.
Identity and Version
- Repository
- CodeMasterCody3D/ternary-composer-charlm
- Publisher
- Cody Dixon
- Task
- Not stated by the source
- Modality
- Other
- Library
- pytorch
- Parameters
- Not stated by the source
- Languages
- Not stated by the source
- Revision
- 673fe99e30cd24d38f8fff9ea248778fd840895f
- First published
- 2026-09-24
- Last updated
- 2026-09-27
Files and Weights
65 files, 5.7 GB in total. The weights are 14 files totalling 5.7 GB in pt.
Every file
| File | Type | Size | SHA-256 |
|---|---|---|---|
| adamA/best_model.pt | Weights | 302.6 MB | 1613a85045ec |
| best_model.pt | Weights | 5.4 MB | b9aab4ab0005 |
| big1/best_model.pt | Weights | 336.3 MB | 5dda49156f84 |
| big1/best_step46000_1.0597.pt | Weights | 336.3 MB | c776716626cb |
| big1/latest.pt | Weights | 1.0 GB | 238dc355cc75 |
| big1chat/best_model.pt | Weights | 336.2 MB | b85d689eb1e8 |
| big1rig/best_model.pt | Weights | 302.6 MB | b057d8af6b77 |
| big1rig/latest.pt | Weights | 907.9 MB | 2ceca2a33f72 |
| big1rig/latest_int8_step3350.pt | Weights | 907.9 MB | 8f377d413871 |
| big1rig/latest_k1_step5200.pt | Weights | 907.9 MB | f178aa862660 |
| ctx16k/best_model.pt | Weights | 5.4 MB | 43802f4da896 |
| ctx2048/best_model.pt | Weights | 5.4 MB | c9a2bf3dea4a |
| flipB/best_model.pt | Weights | 302.6 MB | 3ad2a13da654 |
| live1/best_model.pt | Weights | 5.4 MB | e8518d127e4b |
| bench_rig.py | Configuration | 3.1 KB | — |
| bench_switch_int.py | Configuration | 3.1 KB | — |
| big1/control.json | Configuration | 1.2 KB | — |
| big1rig/control.json | Configuration | 1.1 KB | — |
| charset.json | Configuration | 1.4 KB | — |
| chat.py | Configuration | 13.0 KB | — |
| chat_test.py | Configuration | 4.8 KB | — |
| corpus_stream.py | Configuration | 23.7 KB | — |
| ctx16k/progress.json | Configuration | 16.4 KB | — |
| ctx2048/progress.json | Configuration | 16.3 KB | — |
| diag_ctx.py | Configuration | 4.3 KB | — |
| diag_nan.py | Configuration | 3.3 KB | — |
| diag_srcs.py | Configuration | 2.5 KB | — |
| diag_step.py | Configuration | 2.1 KB | — |
| eval_compare.py | Configuration | 2.9 KB | — |
| fixed_corpus.py | Configuration | 7.2 KB | — |
| flip_lm.py | Configuration | 8.6 KB | — |
| generate.py | Configuration | 4.7 KB | — |
| int_forward.py | Configuration | 15.6 KB | — |
| int_ops.py | Configuration | 8.5 KB | — |
| make_chat_corpus.py | Configuration | 8.6 KB | — |
| smoke_k1.py | Configuration | 2.4 KB | — |
| tern_hip.py | Configuration | 13.1 KB | — |
| tern_kernels.py | Configuration | 8.9 KB | — |
| ternary_lm.py | Configuration | 24.9 KB | — |
| train_best.py | Configuration | 4.5 KB | — |
| train_ctx.py | Configuration | 11.5 KB | — |
| train_live.py | Configuration | 36.5 KB | — |
| README.md | Documentation | 7.9 KB | — |
| adamA/events.jsonl | Other | 25.5 KB | — |
| adamA/train.log | Other | 7.1 KB | — |
| big1/events.jsonl | Other | 171.7 KB | — |
| big1/train.log | Other | 39.7 KB | — |
| big1chat/events.jsonl | Other | 2.3 KB | — |
| big1rig/events.jsonl | Other | 183.8 KB | — |
| csrc/tern_kernels.cu | Other | 24.7 KB | — |
| ctx16k/log.txt | Other | 9.6 KB | — |
| ctx2048/log.txt | Other | 9.2 KB | — |
| flipB/events.jsonl | Other | 26.1 KB | — |
| flipB/train.log | Other | 7.6 KB | — |
| launch_big1.sh | Other | 914 B | — |
| launch_rig.sh | Other | 2.1 KB | — |
| launch_rig_chat.sh | Other | 2.4 KB | — |
| live1/events.jsonl | Other | 188.9 KB | — |
| relaunch_big1.sh | Other | 1.8 KB | — |
| rollover_big1.sh | Other | 2.3 KB | — |
| run_live.sh | Other | 1.1 KB | — |
| .gitattributes | Repository | 1.5 KB | — |
| make_tokenizer.py | Tokenizer | 4.5 KB | — |
| tokenizer.json | Tokenizer | 2.3 MB | — |
| tokenizer.json.stats.json | Tokenizer | 1.8 KB | — |
License and Download
- License
- Not stated by the source
- Access
- Open weights, no gate
- Download size
- 5.7 GB
Released by Cody Dixon through its official repository on Hugging Face.
Memory Requirements
| Precision | Weights in memory |
|---|---|
| As published | 5.7 GB |
Weights only, from the published parameter count; the key-value cache and runtime add to this.