GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
beingamanforever
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
beingamanforever
Model Tree
Qwen/Qwen3.5-4B-Base
Input Modalities
Output Modalities
Supported Functionality
| Run | MATH-500 pass@1 | MuSR macro | IFEval prompt strict | IFBench prompt loose |
|---|---|---|---|---|
| Base student | 12.50 | 43.75 | 43.75 | 12.50 |
| SFT 75:25 | 12.50 | 47.92 | 37.50 | 12.50 |
| OPD 50:50 | 12.50 | 43.75 | 25.00 | 6.25 |
| OPD 75:25 | 6.25 | 45.83 | 31.25 | 25.00 |
| OPD 100:0 | 12.50 | 41.67 | 31.25 | 6.25 |
| Teacher NF4 | 25.00 | 35.42 | 50.00 | 31.25 |
Values are percentages. No blended score is reported. Each task used a fixed 16-example pilot subset; MuSR is the equal-weight macro average of three 16-example subtasks.
For the primary OPD 75:25 versus SFT 75:25 comparison, paired changes were
-6.25 points on MATH-500, -2.08 on MuSR, -6.25 on IFEval, and +12.50 on
IFBench. The 95% paired bootstrap intervals are included in
results/primary_comparison.csv.
The primary adapter was trained for 25 optimizer steps on 40 prompts with a 75:25 reasoning-to-instruction mixture. All OPD runs started from their matching mixture-specific SFT adapter. The teacher was loaded locally with NF4 double quantization and BF16 compute; no remote teacher API was used.
The 128-token training rollout cap truncated 100%, 92%, and 96% of OPD completions for the 50:50, 75:25, and 100:0 runs. Evaluation was also cap-conditioned: 38 to 47 of 48 generative cases per row reached 256 tokens without EOS. These are exploratory systems measurements, not a model-quality estimate.
python
from peft import PeftModelfrom transformers import AutoModelForCausalLM, AutoTokenizeradapter_id = "beingamanforever/Qwen3.5-4B-Whitebox-OPD-Pilot"base_id = "Qwen/Qwen3.5-4B-Base"tokenizer = AutoTokenizer.from_pretrained(adapter_id)base = AutoModelForCausalLM.from_pretrained(base_id, device_map="auto")model = PeftModel.from_pretrained(base, adapter_id)
The weights are LoRA adapters, not a merged full checkpoint. This preserves the base model license and keeps each experimental arm independently usable.
| Path | Contents |
|---|---|
/ | Primary OPD 75:25 adapter and tokenizer |
adapters/opd-50/ | Balanced 50:50 OPD adapter |
adapters/opd-100/ | Reasoning-only OPD adapter |
baselines/sft-75/ | Matching primary SFT adapter |
figures/ | Verified evaluation, mixture, and training charts |
results/ |
Each adapter includes its measured run_metrics.json, trainer_state.json,
and dataset-derived training_provenance.json. Optimizer checkpoints,
serialized training arguments, raw benchmark prompts, and cached base or
teacher weights are intentionally excluded.
The implementation uses TRL 1.12 IWOPDTrainer with sampled-token reverse KL:
text
distillation_objective=iw_opdlmbda=1.0iw_opd_gamma=0.0reverse_kl_top_1_mode=sampledtemperature=1.0top_p=1.0top_k=0
The student and teacher share the exact 248,077-token vocabulary and Qwen control-token IDs, which permits the teacher to score the student's sampled token IDs without retokenization. The experimental target is the local NF4 teacher distribution, not the original BF16 distribution.
Qwen/Qwen3.5-4B-BaseQwen/Qwen3.8-27B, NF4 with BF16 computeA confirmatory study needs at least the planned seeds 42, 43, and 44, 512-token OPD rollouts, uncapped full benchmark evaluation, and matched continued-SFT and corrected RLOO controls. No superiority claim should be made from this pilot.
The implementation, end-to-end tests, per-example scores, exact evaluation provenance, and vector figures are available in the GitHub repository.
Method and benchmark sources:
The adapters are released under Apache-2.0, matching the declared base-model license. Dataset and benchmark licenses remain governed by their respective sources.
| Aggregate scores, paired intervals, cost records, and cap audit |