Jev directory
CommunityJev Models & Reimplementations
Jev Decisions v2 Model
GhostScientist's 0.8B Jev-style decision model: a full fine-tune of Qwen/Qwen3.5-0.8B on 8,702 rows that answers a state/question/options task with exactly one option letter, at a card-reported 0.80 dev accuracy over 400 samples.
- Reading time
- 2 min
- Published by
- Community
- Author
- GhostScientist
- Last verified
- Oct 2026
Works forJev for Classification
Used inJev for Software & DevTools
- Capabilities
- Self-hostedRuns locally
Highlights
- Full fine-tune of Qwen/Qwen3.5-0.8B on GhostScientist/jev-decisions-v1 (8,702 train rows), answering with one option letter.
- Dev results (n=400, greedy): 0.8000 accuracy versus 0.7525 for v1, valid-letter rate 1.000 on both.
- GPU latency mean 138 ms; llama.cpp Q4_K_M on CPU measured 0.820 accuracy at 4 threads over 200 samples.
- Weakest by-source scores: seed_sst5 0.432 and synthetic_scheduling_decision 0.500 (n=8); strongest seed_banking 0.979.
- Trained with TRL 1.14.1 full-parameter SFT on a10g-small for about 77 minutes; final eval_loss 0.1536.
Watch out
Card reviewed 2026-10-17; the dev accuracies are self-reported with no public board and the by-source figures include very small slices (n=8 for synthetic_scheduling_decision). The model card declares no usable licence (its metadata reads licence: license), and because the chat template defaults to thinking mode the single answer letter can end up in thinking or reasoning_content rather than content, dropping measured accuracy to 0.00 unless thinking is disabled.
Questions about this entry
Short answers drawn from the linked project or page.
A Jev/Tev1-style single-letter decision model from GhostScientist: a full fine-tune of Qwen/Qwen3.5-0.8B on the jev-decisions-v1 dataset (8,702 train rows). Given a structured state, question and options task, it answers with exactly one option letter.
Like Jev it turns a state and a question into a decision rather than generated prose, but it is narrower: it emits a single option letter rather than calibrated probabilities per option, so it matches the typed-question workflow without Jev's probability outputs.
Over 400 dev samples, greedy, max_new_tokens 8: 0.8000 accuracy for v2 versus 0.7525 for v1, with a 1.000 valid-letter rate on both, and mean/p50/p95 latency of 138/129/158 ms on an a10g GPU.
The strongest GPU-eval sources are seed_banking at 0.979, seed_agnews at 0.839 and seed_boolq at 0.804; the weakest are seed_sst5 at 0.432, synthetic_scheduling_decision at 0.500 on only 8 samples, and seed_nli at 0.677.
Full-parameter SFT with TRL 1.14.1 and transformers 5.18.0 in conversational prompt-completion format with loss on the completion only: lr 2e-5, cosine schedule, warmup 0.03, effective batch 8, 2 epochs, max length 2,048, bf16 with gradient checkpointing, about 77 minutes on a10g-small.
Because the model was trained on bare-letter completions while the Qwen3.5 chat template defaults to thinking mode: in llama-server the letter can be routed to reasoning_content with empty content. The card measured accuracy going from 0.00 to 0.82 after passing chat_template_kwargs enable_thinking false.
Yes: GhostScientist/jev-decisions-v2-model-gguf hosts Q4_K_M and Q8_0 GGUF exports with a ready-to-use Modelfile, converted with the --no-mtp flag. Neither this model card nor that repository declares licence terms in its metadata.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
X: Standard One SH (StandardThinking)
LinkedIn: What multimodal decision models actually change, and what they do not.
LinkedInJevDirectory.org
Hacker News: Decision Index 0.2.1 scores open reproductions of Jev on a frozen suite.
jackbravo4 pointsDecision Index 0.2.1 – 70 open reproductions of TypeSafe Jev's Decision ModelRead the thread on Hacker NewsHacker News: Pointer to a DiffusionGemma plus vLLM tutorial for a Jev-shaped classifier.
1 pointHow to build a Jev-style classifier with DiffusionGemma and vLLMRead the thread on Hacker NewsMore Jev resources like this
The original OSS Jev Qwen2.5-0.5B decision checkpoint, trained 2,000 steps at learning rate 2e-05; the 10k sibling card warns that this overwrite tip may be the worse noul-fix variant and recommends sized siblings.
Jev Models & Reimplementations#community#open-models#local
Community
GhostScientist's Jev-inspired single-token decision model: a full SFT of Qwen/Qwen3.5-0.8B that answers a structured state/question/options task with exactly one option letter, at a card-reported 0.755 greedy dev accuracy against 0.657 for the base model.
Jev Models & Reimplementations#community#open-models#local
Community
A local yes/no decision model that speaks Jev's wire format on a laptop CPU, refusing Choice and Score until those question types ship.
Jev Models & Reimplementations#community#python#open-models
Community
feder-cr/jevC++
feder-cr/jevC++A local daemon that pulls open decision models and serves them on TypeSafe's /v1/systemone, so a Jev client switches hosts with TYPESAFE_BASE_URL.
Jev Models & Reimplementations#community#rust#open-models
Community
ollaya-dev/ollayaRust
ollaya-dev/ollayaRustA local server that reads next-token probabilities from a fine-tuned Spark model and returns Choice, Score, and Noul answers with zero generated tokens.
Jev Models & Reimplementations#community#python#open-models
Community
Rizzo-AI-Academy/rizzo-flowPython
Rizzo-AI-Academy/rizzo-flowPythonAn Apache-2.0 decision engine with a TypeSafe-compatible endpoint, document answers with source passages, and a ViZDoom demo that picks actions from structured game state.
Jev Models & Reimplementations#community#python#local
Community
deepanwadhwa/OpenDecisionPython
deepanwadhwa/OpenDecisionPythonFrom the Jev guides
Original write-ups on what this entry is for.
From the Jev AI model community
Posts from builders shipping with Jev right now.