What changed
Same EDSL survey, two respondents (a budget-conscious teacher on 52k,asurgeonon480k):
Table with columns: question, released Kev-4B, this model| question | released Kev-4B | this model |
|---|
| Streaming service at 8vs40/month, teacher: P(subscribe) | 0.49 → 0.25 | 0.88 → 0.05 |
| same, surgeon | 0.78 → 0.77 | 0.96 → 0.51 |
| Picks the dominated plan (45for5GBnextto30 for 5 GB) | 0.24–0.29 | 0.02–0.03 |
| Surgeon "very likely" to buy an air fryer at 8vs40 | 0.25 → 0.35 (wrong way) | 0.82 → 0.67 |
Results
Table with columns: DLS v0.2 (ours, 9 families), blind test (513 contrasts by GPT-6.1-sol), bank-run battery: payoff shift (theory 80), lottery slope (theory 1.0), EDSL-General | DLS v0.2 (ours, 9 families) | blind test (513 contrasts by GPT-6.1-sol) | bank-run battery: payoff shift (theory 80) | lottery slope (theory 1.0) | EDSL-General |
|---|
| this model | 79 [73, 86] | 81 [78, 84] | 28 | 0.51 | 85 |
| Kev-4B (released) | 53 | 67 | −1 | 0.05 | 85 |
| Lory-27B v1 | 88 | 92 | 66 | 0.77 | 88 |
| Perplexity Decider v1.2 | 77 | 93 | 9 | 0.43 | 88 |
Gains over the released model are large and robust (DLS +25.5 [+17.9, +32.9], blind +14.0 [+10.4, +17.5]). It ties
Decider on DLS (+1.6 [−5.2, +8.2]) but is clearly behind it on the blind test (−12.7 [−15.5, −10.1]). Test
descriptions: see the Lory-27B card.
Use
git clone https://github.com/jaredpalmer/kev && cd kev && git checkout 2074cc4 && uv sync --extra serveuv run python -m kev.serve --run eltokh7/lory-4b --device cuda --port 8850 # or --device mps on Apple silicon
Requests as in the Lory-27B card; EDSL surveys via edsl_decision.py (in this repo). Cost: at least 53 requests/s on one
A100 (client-limited), under $0.25 per 1,000 respondents × 20 questions.
Training
The same 27,000 records and recipe as Lory-27B (LoRA delta from the released adapter, + 5,000 Kev decision-v7
replay, one epoch, lr 2e-5, batch 8, bf16), one A100, ~1.8 hours.
Limitations
Everything on the Lory-27B card, plus: it carries the logic to unfamiliar wording much less well than the 27B
(bank-run payoff shift 28 vs 66); the persona signal is slightly weaker than before training on the blind test (69
vs 72); framing invariance 56/100.
Attribution
Kev and Kev-4B by Jared Palmer (Apache-2.0); backbone Qwen3.5-4B-Base (Qwen, Apache-2.0).