Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,21 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
tags: [anima, clm, bytegpt, 7b, lane-g, h1141, 7b-pass-attempt]
|
| 4 |
+
base_model: dancinlab/clm-v1-ref-pytorch-cuda-7b-broad-converged
|
| 5 |
+
---
|
| 6 |
+
# anima CLM 7B PASS-attempt (H_1141)
|
| 7 |
+
|
| 8 |
+
Continue-trained 7B ByteGPT (vocab256/d4096/36L/32H, 7.25B) from the H_1128
|
| 9 |
+
broad-converged 7B on the balanced 5-lang 1.5GB corpus, pushed toward coherence,
|
| 10 |
+
then run through the FULL frozen gate battery in `/7B_PASS_CONDITIONS.md`.
|
| 11 |
+
|
| 12 |
+
**Gate tally:** `G0=P,G1=P,G2=P,G3=P,val=1.1856881439685822`
|
| 13 |
+
**Visibility:** PUBLIC (PUBLIC iff G0∧G1∧G2 all PASS — a_hf_autonomous).
|
| 14 |
+
|
| 15 |
+
- G0 COHERENCE — kwr>=0.50 vs /usr/share/dict on >=4/5 English concept prompts
|
| 16 |
+
- G1 RECOMBINATION — H_1137 metric (English + >=3/5 langs)
|
| 17 |
+
- G2 NOVELTY — H_1140 metric (>=3 corpus-absent coherent n-grams, control=0)
|
| 18 |
+
- G3 PHILOSOPHY — byte-continuation only (p1-p8), held by construction
|
| 19 |
+
|
| 20 |
+
ckpt sha256: `4de903714112c26c826a983797e5dfea0c7b3c1f19f15a34dd35822f33e245d9`. Lane-G PyTorch-CUDA REFERENCE mouth (a_clm_gen_pipeline),
|
| 21 |
+
NOT the CORE substrate. a_scale_honest_scope: single 7B rung, surface-level metrics.
|