Commit History
17 days ago
-
document the unsloth-free loader for this arm
Verified signature42ba04f
18 days ago
-
document the unsloth-free inference path (m15)
Verified signature7835600 -
install: klev is not on PyPI - install from GitHub
Verified signature504b4a9 -
Usage: use the installed package API and console scripts
Verified signature5375c83 -
Install section: pip install klev
Verified signatureff3fb1b -
Usage: validated in-process and HTTP paths
Verified signatureklev is a library first, so the card now leads with an in-process call rather than a server. model/infer.py is the shared core: it reuses data/encoding.py, data/format.py and model/head.py with the scorer rather than reimplementing them, so the probabilities a user sees are the ones eval/eval_decisions.py measures. Plus examples/system_one.py for the shell and serve/server.py for when HTTP is actually wanted. Every snippet was executed. Real output is pasted, not illustrative: noul p(true)=0.9169, choice atm 0.6136, score 1.485, rejection 0.168 and 0.027. The temperature table is called out because head.pt stores 1.0 (written during training; the head only tempers in eval mode) and serving raw costs 0.08 ECE. The server routes match kev's own shapes, so eval/eval_systemone_http.py scores it unmodified -- verified over 120 decision-v7 development rows, zero failures. Also records that the ~195-token ceiling on the RX 6600M is a training limit only: a 703-token row was served 200 with …d0bef41 -
Document base/ in the card
Verified signaturebd07490 -
Add base/: the same recipe on Qwen/Q.5-0.8B-Base, kev's checkpoint
Verified signatureHolds the base fixed against kev-0.8B. The original comparison used different base checkpoints, and the base sets how strong an external task LoRA can be (tweet_eval probe 0.715 on the IT base vs 0.660 on -Base, same data and hyper-parameters), so the stitch gap measured the adapter as much as the decision model. base/ contains its own adapter, head.pt, train_config.json, the external LoRA that is SHARED with kev-0.8B, and the fitted 32-shot probe. On that shared adapter the stitch gain is +7.5 pts for klev against +3.0 for kev, so roughly half the original +8.0 gap was the adapter's base and half is the recipe. 195.2 min to train on an RX 6600M. Teacher cache rebuilt (1,413,968 positions) because the KL anchor is computed from the frozen base's own logits. Decision numbers base-matched: dev 0.8065 (kev 0.8120), transfer 0.6283 (kev 0.6126).6d82b31 -
klev-0.8b: Qwen3.5-0.8B decision model (QLoRA + pointer head + rejection channel)
Verified signatureQLoRA adapter (r=16, alpha=32) and pointer head for Qwen/Qwen3.5-0.8B, trained on decision-v7 for 2 epochs (3,894 steps) with a KL anchor 0.3 to the frozen base. 0.89 GB in 4-bit, 1.39 GB peak; 192 min on an 8 GB RX 6600M (gfx1032, ROCm 7). decision-v7 dev 0.8120 (ties kev-0.8B on identical rows), transfer-v4 dev 0.6636 (+5.1 pts over kev-0.8B), each served with its own fitted temperature. Includes the few-shot stitch: two external task LoRAs (tweet_eval hate, RumourEval 2019) with fitted 32-shot nearest-class-mean probes. On TweetEval the fused klev equals the external LoRA's own probe (0.715 vs 0.605 for the bare head), recovering 100 pct of the available margin from 32 labelled rows and two scalars.5901eb4