GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
aziz9788
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
aziz9788
Model Tree
unsloth/Qwen3.5-4B
Input Modalities
Output Modalities
Supported Functionality
| Field | Value |
|---|---|
| Stage | select-release (completed) |
| Decision | reject_opd_keep_parent |
| Selected release | locked parent e11_dpo_v3 (merged E11 + DPO-v3) |
| OPD child | rejected (IFEval non-regression fail) |
| RLVR | skipped / cancelled (diagnose gate unmet) |
| Restorative DPO | recommended, not started |
Non-regression rule: no task loss >1pp vs locked parent; Math+Pro prefer improvement. IFEval instruction-strict lost −4.78pp → gate FAIL.
Protocol: local AraMath / AraPro / AraIFEval, seed 42, enable_thinking=false, TP=2.
| Task | OPD (this adapter) | Floor (parent E11+DPO-v3) | Δ (pp) | Gate |
|---|---|---|---|---|
| AraMath | 72.40 | 68.60 | +3.80 | ok |
| AraPro | 64.93 | 63.93 | +1.00 | ok |
| IFEval inst-strict | 77.20 | 81.98 | −4.78 |
STEM OPD improved Math/Pro but erased IF calibration (STEM-only accepted set; no_if=true). That fails the preregistered gate.
PEFT LoRA adapter only (rsLoRA r=32, α=32, 1 epoch, LR 2e-6, completion-only SFT).
| Item | Detail |
|---|---|
| Train mode | prompt/completion SFT (completion_only_loss) — not DPO under OPD |
n_train | 445 STEM pairs |
| Teachers | primary z-ai/glm-5.2; fallback deepseek/deepseek-v4-flash-0731 |
| IF in OPD data | none (no_if=true) |
| Train loss | ~1.82 |
| Adapter SHA256 |
Fail-closed checks at select-release:
merged_sft_e11_fixed), not raw Qwen4957d013…); OPD path ≠ DPO pathmerged_sft_e11_fixed + opd_stem_sft_child_1ep (no DPO under OPD)Do not load this adapter on raw unsloth/Qwen3.5-4B alone and expect the reported scores.
markdown
unsloth/Qwen3.5-4B @ 3764fa359b9082ea5a1e4a5e3ac3aaf6e9671636│▼merged E11 SFT (Vast: .../checkpoints/merged_sft_e11_fixed)│├──► + DPO-v3 LoRA → LOCKED PARENT / campaign release ✅│└──► + OPD STEM LoRA → THIS REPO (rejected) ❌(trained on merged E11 alone; DPO not under OPD)
Base (public): unsloth/Qwen3.5-4B revision 3764fa359b9082ea5a1e4a5e3ac3aaf6e9671636.
Runtime parent for this LoRA: the E11 merged BF16 checkpoint used in training (base_model_name_or_path in adapter_config.json still points at the Vast merge path). Full BF16 merge is not uploaded here (disk / size). To reproduce the evaluated OPD stack you need that merged E11 parent (or an equivalent merge of the E11 campaign SFT), then:
python
from peft import PeftModelfrom transformers import AutoModelForCausalLM, AutoTokenizerbase = AutoModelForCausalLM.from_pretrained("<path-or-repo-to-merged-e11>", # NOT raw Qwen for reported scorestorch_dtype="bfloat16",device_map="auto",)model = PeftModel.from_pretrained(base, "aziz9788/qwen35-4b-opd-stem-sft-child")
OPD was STEM-only distillation (445 accepted pairs; no IF). Math/Pro rose, IFEval fell past the 1pp cap. Campaign keeps merged E11 + DPO-v3. Restorative DPO is the suggested next step to recover IF while trying to keep STEM gains — not started in this release.
| Path | Description |
|---|---|
adapter_model.safetensors / adapter_config.json | PEFT LoRA weights |
tokenizer* / chat_template.jinja | tokenizer artifacts from training export |
select_release_report.md | human-readable select-release report |
select_release_scores.json | machine-readable decision + scores |
select_release/ |
LIMIT=64 checkpoint sentinels (ckpt-14/28) are smoke-only and not comparable to full-suite floors; they were not used for promotion.
| IFEval prompt-strict | 50.75 | 58.96 (ref) | −8.21 | FAIL |
b22f733da818018b142d24cb8eeefac4ef6d6a05875565ad8e1e6bcac4176668 (adapter_model.safetensors) |
| training record, adapter config snapshot, eval summaries |
OPD_STATUS.md | stage status (OPD SFT + select-release complete) |