GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
gyanbhambhani
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
gyanbhambhani
Model Tree
Qwen/Qwen2.5-0.5B-Instruct
Input Modalities
Output Modalities
Supported Functionality
Answer:\n (train/infer lockstep). Binary head
weights (noul_binary_head.pt) are on Hub for ablation only
(OSS_JEV_NOUL_BINARY_HEAD=1).Not intended for: production Noul gates that must beat a constant-N prior, safety-critical yes/no, or claiming parity with live Jev.
| Split | Frozen 98-row prior distill eval |
| Metric | Argmax agree vs teacher hard labels; Noul @ P(Y)≥0.5 |
| Prompt | Answer:\n + train-aligned codes |
| Bias | 0 (raw LM argmax; not “argmax via bias”) |
| N | 28 Choice / 43 Noul / 27 Score |
| Primary surface | LM Y/N (not binary head) |
Treat ±~10pp overall as noise (bootstrap on N=98).
| Metric | Value |
|---|---|
| Overall | 60.2% |
| Choice | 82.1% (23/28) |
| Noul | 44.2% (19/43) — 23 Y / 20 N |
| Score level | 63.0% (17/27) |
| Score Pearson r | 0.768 |
| P(Y) mean / min / max | 0.487 / 0.083 / 0.688 |
Train: continue 2000 steps from noul-aug, train_loss≈0.6747.
Job 6ad1e316c66288145225ac74 · a10g-small · ~15.7 min.
Hub sha fd6bca665e9f91ab9e96d4c4f43a5d59b3c5b19e.
| Condition | Overall | Choice | Noul | Noul preds | Score lvl | Score r |
|---|---|---|---|---|---|---|
| 10k baseline | 45.9% | 82.1% | 16.3% | 43/43 Y | 55.6% | 0.804 |
| noul-aug | ~46.9% | 85.7% | 16.3% | 43/43 Y | 55.6% |
| Criterion | Result |
|---|---|
| Noul not all-Y | PASS (20/43 N via LM) |
| Choice ≥ ~80% | PASS (82.1%) |
| Beat constant-N | FAIL (44.2% ≪ 83.7%) |
| Not “argmax via bias” | PASS (bias=0) |
Answer:\n so Y/N are single trailing tokens.Do not default to the binary head for decisions — use LM Y/N.
Optional: load noul_binary_head.pt only for ablation.
python
# Pseudocode — LM Y/N serving (primary)codes, keys = resolve_codes_and_keys(question)# prompt must end with "Answer:\\n"logits = model(...).logits[0, -1, code_token_ids]probs = softmax(logits)pred = codes[probs.argmax()]Default noul_logit_bias_n=0.
Answer:\n lockstepWithout the newline, Qwen2.5 merges Answer:N → :N while
Answer:Y splits cleanly — historically locked all-Y. Train and
infer must both use Answer:\n.
The Linear(h→2) aux CE moved the LM surface off all-Y. Serving the head itself stays 43/43 Y. Keep head for training; serve LM.
44.2% Noul beats the 16.3% all-Y floor and edges ZS-1.5B Noul (41.9%), but constant-N (83.7%) remains stronger. Prefer constant-N or live Jev when Noul competence is the product bar.
N=98 → wide CIs. Do not overclaim small pp gaps.
| Continue-from | gyanbhambhani/oss-jev-qwen05-decision-noul-aug |
| Recipe | soft_ce + constrained_codes + balance_noul (n_frac=0.5) |
| Augments | noul_random_codes + negation_pairs + cons_w=1.0 |
| NEW | noul_binary_head=True (weight 1.0) as train aux |
| Bias | noul_logit_bias_n=0.0 |
| Train-src | gyanbhambhani/oss-jev-train-src |
Never overwrote …-decision / …-10k / …-noul-aug.
text
@misc{oss-jev-qwen05-decision-noul-bin, title={OSS Jev Noul binary-head continue (0.5B)}, author={Bhambhani, Gyan}, year={2026}, howpublished={Hugging Face}, url={https://huggingface.co/gyanbhambhani/oss-jev-qwen05-decision-noul-bin}}| 0.774 |
| noul-bin LM Y/N | 60.2% | 82.1% | 44.2% | 23 Y / 20 N | 63.0% | 0.768 |
| noul-bin binary head | 48.0% | 82.1% | 16.3% | 43/43 Y | 63.0% | — |
| Constant-N | — | — | 83.7% | 43/43 N | — | — |
| Hardware | HF Jobs a10g-small, bf16, 2000 steps, lr 1e-5 |
| Config | configs/jobs-a10g-noul-bin.yaml |