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