4.87 GB

Ctrl+K

1 contributor

History: 251 commits

CodeMasterCody3D

big1: best_model.pt, events.jsonl

79cc9b9 verified less than a minute ago

adamA
adamA: best_model.pt, events.jsonl, train.log about 18 hours ago
big1
big1: best_model.pt, events.jsonl less than a minute ago
big1rig
big1: rig state at step ~7850 (best 1.182) handed back to the T4; control for batch 8 about 4 hours ago
csrc
code: base-3 (5 trits/byte) container + k1 KV cache kernels, chat.py (KV-cached decode), docs about 12 hours ago
ctx16k
ctx16k final (negative result) + live1 final (2.052 b/c, 12-source mix) + 32k BPE tokenizer + v2 code (int8 emb/KV, res shift, final norm, packing) + width-probe results in README 1 day ago
ctx2048
ctx2048 final (1.219 b/c) + ctx16k (in progress) + live-mode/v2 code (streaming mix, int8 emb/KV, BPE) + README 1 day ago
flipB
flipB: best_model.pt, events.jsonl, train.log about 19 hours ago
live1
ctx16k final (negative result) + live1 final (2.052 b/c, 12-source mix) + 32k BPE tokenizer + v2 code (int8 emb/KV, res shift, final norm, packing) + width-probe results in README 1 day ago
.gitattributes

1.52 kB
initial commit 1 day ago
README.md

7.85 kB
flip_lm.py (low-rank-Adam trit flips, no fp latents) + fixed_corpus.py + train_live --flip/--fixed-corpus; rollover/relaunch scripts; README: big1 + flip mode about 21 hours ago
bench_rig.py

3.14 kB
ROCm-safe flip rule (kthvalue/where), --accum, --freeze-aux, --lr-horizon, launch_rig.sh/bench_rig.py for the RX 5600 XT rig about 21 hours ago
bench_switch_int.py

3.13 kB
code: HIP kernels (csrc/tern_kernels.cu, tern_hip.py), integer path fix (int32 residual, int16 GELU), docs about 15 hours ago
best_model.pt

5.44 MB
xet

Bank strict-ternary composer char-LM (0.723 acc / 1.305 bpc, alive-init + 4-lane composer) 1 day ago
charset.json

1.36 kB
Bank strict-ternary composer char-LM (0.723 acc / 1.305 bpc, alive-init + 4-lane composer) 1 day ago
chat.py

10.1 kB
code: base-3 (5 trits/byte) container + k1 KV cache kernels, chat.py (KV-cached decode), docs about 12 hours ago
corpus_stream.py

22.4 kB
ctx2048 final (1.219 b/c) + ctx16k (in progress) + live-mode/v2 code (streaming mix, int8 emb/KV, BPE) + README 1 day ago
diag_ctx.py

4.34 kB
QK-norm (bounded ternary attention scores) + diag_ctx; big1 trajectory 2.88 -> 1.87 b/c in README-adjacent docs 1 day ago
diag_nan.py

3.27 kB
v2 fixes: folded residual shift (fp16-safe lanes), fp32 int8 attention, scaler/Adam validation on resume, grad clip, NaN-storm exit; diagnostics; README/probe results 1 day ago
diag_srcs.py

2.47 kB
v2 fixes: folded residual shift (fp16-safe lanes), fp32 int8 attention, scaler/Adam validation on resume, grad clip, NaN-storm exit; diagnostics; README/probe results 1 day ago
diag_step.py

2.08 kB
v2 fixes: folded residual shift (fp16-safe lanes), fp32 int8 attention, scaler/Adam validation on resume, grad clip, NaN-storm exit; diagnostics; README/probe results 1 day ago
eval_compare.py

2.87 kB
ctx2048 final (1.219 b/c) + ctx16k (in progress) + live-mode/v2 code (streaming mix, int8 emb/KV, BPE) + README 1 day ago
fixed_corpus.py

7.19 kB
flip_lm.py (low-rank-Adam trit flips, no fp latents) + fixed_corpus.py + train_live --flip/--fixed-corpus; rollover/relaunch scripts; README: big1 + flip mode about 21 hours ago
flip_lm.py

8.62 kB
ROCm-safe flip rule (kthvalue/where), --accum, --freeze-aux, --lr-horizon, launch_rig.sh/bench_rig.py for the RX 5600 XT rig about 21 hours ago
generate.py

4.67 kB
ctx2048 final (1.219 b/c) + ctx16k (in progress) + live-mode/v2 code (streaming mix, int8 emb/KV, BPE) + README 1 day ago
int_forward.py

15.6 kB
code: base-3 (5 trits/byte) container + k1 KV cache kernels, chat.py (KV-cached decode), docs about 12 hours ago
int_ops.py

8.49 kB
code: HIP kernels (csrc/tern_kernels.cu, tern_hip.py), integer path fix (int32 residual, int16 GELU), docs about 15 hours ago
launch_big1.sh

914 Bytes
QK-norm (bounded ternary attention scores) + diag_ctx; big1 trajectory 2.88 -> 1.87 b/c in README-adjacent docs 1 day ago
launch_rig.sh

2.15 kB
composer switch (adaptive 1..N lane routing, live-toggleable), short conv (zero-init), all-integer inference path (int_ops/int_forward), 2-bit packed GEMV (Triton on NVIDIA) + switch decode about 15 hours ago
make_tokenizer.py

4.52 kB
ctx16k final (negative result) + live1 final (2.052 b/c, 12-source mix) + 32k BPE tokenizer + v2 code (int8 emb/KV, res shift, final norm, packing) + width-probe results in README 1 day ago
relaunch_big1.sh

1.75 kB
flip_lm.py (low-rank-Adam trit flips, no fp latents) + fixed_corpus.py + train_live --flip/--fixed-corpus; rollover/relaunch scripts; README: big1 + flip mode about 21 hours ago
rollover_big1.sh

2.1 kB
flip_lm.py (low-rank-Adam trit flips, no fp latents) + fixed_corpus.py + train_live --flip/--fixed-corpus; rollover/relaunch scripts; README: big1 + flip mode about 21 hours ago
run_live.sh

1.08 kB
ctx2048 final (1.219 b/c) + ctx16k (in progress) + live-mode/v2 code (streaming mix, int8 emb/KV, BPE) + README 1 day ago
smoke_k1.py

2.38 kB
code: base-3 (5 trits/byte) container + k1 KV cache kernels, chat.py (KV-cached decode), docs about 12 hours ago
tern_hip.py

13.1 kB
code: base-3 (5 trits/byte) container + k1 KV cache kernels, chat.py (KV-cached decode), docs about 12 hours ago
tern_kernels.py

8.93 kB
code: base-3 (5 trits/byte) container + k1 KV cache kernels, chat.py (KV-cached decode), docs about 12 hours ago
ternary_lm.py

24.9 kB
code: base-3 (5 trits/byte) container + k1 KV cache kernels, chat.py (KV-cached decode), docs about 12 hours ago
tokenizer.json

2.3 MB
ctx16k final (negative result) + live1 final (2.052 b/c, 12-source mix) + 32k BPE tokenizer + v2 code (int8 emb/KV, res shift, final norm, packing) + width-probe results in README 1 day ago
tokenizer.json.stats.json

1.76 kB
ctx16k final (negative result) + live1 final (2.052 b/c, 12-source mix) + 32k BPE tokenizer + v2 code (int8 emb/KV, res shift, final norm, packing) + width-probe results in README 1 day ago
train_best.py

4.48 kB
Bank strict-ternary composer char-LM (0.723 acc / 1.305 bpc, alive-init + 4-lane composer) 1 day ago
train_ctx.py

11.5 kB
ctx2048 final (1.219 b/c) + ctx16k (in progress) + live-mode/v2 code (streaming mix, int8 emb/KV, BPE) + README 1 day ago
train_live.py

31.7 kB
code: k1 KV-mode resume fix, chat.py KV header; docs (k1 episode, T4 handover) about 4 hours ago