GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
Prience91
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
Prience91
Model Tree
Qwen/Qwen3.8-27B
Input Modalities
Output Modalities
Supported Functionality
Checkpoint 230 (epoch 1.0) achieved validation loss 0.3242049217224121, the lowest among evaluations completed through step 280 at export selection.
Base: Qwen/Qwen3.8-27B, pinned revision 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0.
Dataset: Prience91/Climate-1000-canary, pinned revision 18d438de908614cbe34ffcd91a563b3e8e320cc9.
3,665 training examples and 406 validation examples. All 100 reserved held-out paper identities were excluded from SFT. The 78 approved test references remain provisional. No canonical benchmark or deployment-quality claim is made.
QLoRA with frozen NF4 4-bit base, double quantization and BF16 compute. LoRA rank 32, alpha 64, dropout 0.05, all supported linear projections. Saved adapter tensors are FP32, as verified directly from the safetensors header; BF16 compute does not establish adapter parameter storage dtype.
Completion-only loss; thinking disabled; maximum length 8,192; no packing. Microbatch 1, accumulation 16, learning rate 1e-4, seed 3407. The full run targets two epochs / 460 optimizer steps with evaluation every ten optimizer steps.
Recorded environment: PyTorch 2.11.0, Transformers 5.15.1, TRL 0.24.0, PEFT 0.19.1, bitsandbytes 0.50.0, datasets 4.3.0.
Load the pinned base revision in a compatible Transformers/PEFT environment, then attach this adapter with PeftModel.from_pretrained. Match the training NF4, double-quantization and BF16-compute settings when reproducing its quantized base representation. Apply the saved chat template with enable_thinking=False.
The adapter configuration uses the public base identifier and pinned revision instead of a Spark-local cache path. The tokenizer and chat template accompany the adapter. Optimizer, scheduler, RNG and serialized training arguments are excluded: this package is for adapter loading, not full training resumption. The complete original checkpoint remains archived on Spark.
The project later reloads the pinned base in BF16, merges the selected adapter, and saves a separate BF16 artifact for vLLM. This publication performs no merge, inference activation, or post-merge evaluation.