openjev — Qwen3.5 trained as jev model

openjev is Qwen3.5 turned into a jev model: a single cross-encoder that reads a premise and a hypothesis and answers with entailment, contradiction or neutral. That one primitive is enough to rerank answers, grade them against a reference, guard content, and play games in real time: hand it the game state and a few statements about it, and the argmax entailment is the move. Nothing is trained per task.

openjev-4B v2: text, images and agents

The new 4B checkpoint (qwen3.5-4b-nli-v2/) reads images as well as text and was trained on a much larger and harder mixture. It is strictly zero-shot on everything shown here.

• Doom straight from the pixels (first video): 10.4 kills per episode, twice the v1 model (5.2); random play gets 1.
• Crafts an iron pickaxe from nothing in real Minecraft (second video): 11 milestones in ~22 decisions, driven by a backward-chaining scaffold where the jev model only checks statements about the inventory and the world.
• Much stronger on adversarial NLI (ANLI r3 0.42 → 0.63, WANLI 0.63 → 0.77) and on image claims (0.52 → 0.84), better reranking (ARC-Challenge 0.59 → 0.72, MMLU 0.47 → 0.53), same MNLI (0.91).

Doom from the text state (v2, 11 kills per episode; a perfect-information bot gets 18.8):

Bigger jev: Qwen3.5-35B-A3B (MoE) as the backbone (qwen3.5-35b-a3b-nli/). Zero-shot, and with the backbone frozen plus a small MLP head on the last-token latent (mlp_heads_35b/, one head per task, loadable with LatentMLPHead.load):

What's inside

• qwen3.5-0.8b-nli-v2s-long/ — the small v2s checkpoint (0.8B, 4k context): the v2 mixture plus faithfulness / instruction-following / false-premise data and a long-document stage. MNLI 86.2/87.1, ANLI r1 65.1, SciTail 93.2, RAGTruth AUROC 0.90, LLM-AggreFact avg bAcc 69.5. This is what the demo Space serves.
• qwen3.5-4b-nli-v2/ — recommended: the 4B v2 jev checkpoint, text + images.
• qwen3.5-4b-nli/ — the original 4B jev checkpoint (text).
• qwen3.5-35b-a3b-nli/ — the 35B-A3B MoE jev checkpoint (load with modeling_qwen35_moe_seqcls.py).
• All checkpoints: Qwen3_5ForSequenceClassification, 3 labels contradiction, entailment, neutral, last-token pooling, trained with plain cross-entropy over the three classes.
• modeling_openjev.py — OpenJevCrossEncoder: predict, predict_hypotheses, rerank, grade, latents, latents_hypotheses; LatentMLPHead for the per-task heads.
• modeling_qwen35_moe_seqcls.py — Qwen3_5MoeForSequenceClassification for the 35B-A3B backbone.
• mlp_heads_35b/<task>/ — head.pt + norm.npz + meta.json, the 35B latent + MLP heads behind the second radar.
• code/ — everything used here: the trainer and data mixture builder, the evaluation harness, Flappy Bird, Doom (text and pixels), the Minecraft scaffold and bot, the radar, the SGLang package / launcher / client / benchmark.
• videos/ — Flappy Bird, Doom and Minecraft replays; results/ — raw JSON for every run and the full report.

Use it

from modeling_openjev import OpenJevCrossEncoder
jev = OpenJevCrossEncoder("AlexWortega/openjev", subfolder="qwen3.5-4b-nli-v2")

jev.predict([("The bird is 0.05 below the centre of the gap.", "The bird is below the centre of the gap.")])
# -> [[contradiction, entailment, neutral]] probabilities

jev.rerank("Which gas do plants absorb during photosynthesis?", ["oxygen", "carbon dioxide", "nitrogen"])
# -> index of the option with the highest entailment

jev.predict_hypotheses("Which gas do plants absorb during photosynthesis?",
                       ["The correct answer is: oxygen", "The correct answer is: carbon dioxide",
                        "The correct answer is: nitrogen"])
# -> one [contradiction, entailment, neutral] row per hypothesis

predict_hypotheses and latents_hypotheses use the existing pairwise batch for one or two hypotheses. For three or more, they compute the common token prefix once, then score every hypothesis in one batched continuation. rerank uses the same rule. Qwen3.5 has recurrent linear-attention layers, so a 4D packed tree mask alone would mix branches; the shared prefix cache is copied into separate batch entries for the suffixes. This path is for text inputs. Its batch size and suffix padding use memory proportional to the number and length of the hypotheses; split very large option sets into smaller calls.

Or with plain transformers:

from transformers import AutoModelForSequenceClassification, AutoTokenizer
tok = AutoTokenizer.from_pretrained("AlexWortega/openjev", subfolder="qwen3.5-4b-nli-v2")
model = AutoModelForSequenceClassification.from_pretrained("AlexWortega/openjev", subfolder="qwen3.5-4b-nli-v2")
text = model.config.nli_template.format(premise="...", hypothesis="...")

Images go inside the premise as <|vision_start|><|image_pad|>…<|vision_end|> with pixel_values / image_grid_thw from the Qwen3.5 image processor; see code/doom_vision.py and code/eval_image_nli.py.

Serve it with SGLang

SGLang has no sequence-classification class for Qwen3.5, so code/sglang_openjev/ is an external model package that adds the score head to SGLang's Qwen3_5ForConditionalGeneration (hybrid cache, mrope and the vision tower stay as they are). Text and images both work; /classify returns the three raw logits.

hf download AlexWortega/openjev --include "qwen3.5-0.8b-nli-v2s-long/*" "code/*" --local-dir openjev
cd openjev/code && bash serve_sglang.sh ../qwen3.5-0.8b-nli-v2s-long 30000      # tested with sglang 0.5.19

from sglang_client import OpenJevSGLang          # code/sglang_client.py
jev = OpenJevSGLang("http://127.0.0.1:30000")
jev.predict([("A man is playing a guitar.", "Someone is making music.")])    # [[con, ent, neu]]
jev.predict([("A photograph of a scene:", "There is a dog.")], images=["dog.jpg"])

Without the script: SGLANG_EXTERNAL_MODEL_PACKAGE=sglang_openjev with code/ on PYTHONPATH, then python -m sglang.launch_server --model-path <dir> --is-embedding --json-model-override-args '{"architectures": ["Qwen3_5ForConditionalGeneration"]}'.

Same predictions as transformers, 1.5-3x the throughput. qwen3.5-0.8b-nli-v2s-long on one RTX A6000 that was shared with another job (so absolute speed is a lower bound), transformers at batch 32 (code/bench_sglang.py):

task	pairs	tokens / pair	acc SGLang	acc transformers	SGLang pairs/s	transformers pairs/s	speed-up
MNLI m+mm	19647	44	86.63	86.64	321	152	2.1x
ANLI r1	1000	105	65.3	64.9	189	119	1.6x
ANLI r2	1000	103	50.6	50.8	181	110	1.6x
ANLI r3	1200	92	48.6	48.6	218	95	2.3x
WANLI	5000	40	73.7	73.7	400	196	2.0x
SciTail	2126	43	93.3	93.4	319	217	1.5x
ConTRoL (long)	805	612	51.9	51.4	45	15	3.1x
ARC-Challenge (rerank)	9374	49	49.2	49.2	334	197	1.7x
HellaSwag (rerank, 2k questions)	16000	114	37.6	37.5	176	109	1.6x

Reference point: dleemiller's NLI cross-encoders. Licence MIT.