GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
wli14
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
wli14
Model Tree
Qwen/Qwen3-8B
Input Modalities
Output Modalities
Supported Functionality
Qwen/Qwen3-8Br=16, alpha=32, dropout 0.05q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_projThe source corpus contained 3,644 six-turn evidence-reasoning dialogues, or 21,864 valid user-assistant turns. Each turn-level training item included a selected-spot evidence card, the current question, up to two preceding dialogue pairs and one assistant target.
A deterministic selection with seed 42 produced:
The raw training records are not included in this model repository. Aggregate provenance and split statistics are provided in training/data_manifest.json.
Training used four NVIDIA H100 80-GB GPUs for one epoch with BF16 weights and FlashAttention 2. Only assistant target tokens contributed to the loss.
| Setting | Value |
|---|---|
| Optimizer steps | 625 |
| Global batch size | 16 |
| Maximum sequence length | 2,304 |
| Learning rate | 2e-4 |
| Weight decay | 0.01 |
| Warm-up ratio | 0.03 |
| Schedule | Cosine |
| Validation loss | 0.6345 |
| Training and validation time | 1,056.9 s |
The exact training summary, scripts and data manifest are included in this repository.
python
import jsonimport torchfrom peft import PeftModelfrom transformers import AutoModelForCausalLM, AutoTokenizer base_model_id = "Qwen/Qwen3-8B"adapter_id = "wli14/HistAgent-Qwen3-8B-LoRA-10K" tokenizer = AutoTokenizer.from_pretrained(adapter_id)base_model = AutoModelForCausalLM.from_pretrained( base_model_id, torch_dtype=torch.bfloat16, device_map="auto",)model = PeftModel.from_pretrained(base_model, adapter_id).eval() evidence_card = { "spot": {"species": "human", "organ": "kidney"}, "ranked_genes": ["GENE1", "GENE2"], "cell_type_composition": [], "pathway_evidence": {}, "spatial_context": {}, "quality_flags": {},}question = "What tissue state is supported by this spot?" messages = [ { "role": "system", "content": ( "You are HistAgent, a molecular reasoning assistant for histology images. " "Use only the supplied evidence card to answer questions about the selected " "tissue spot. Do not fabricate genes, pathways, cell types or spatial " "conclusions that are absent from the evidence. State uncertainty when the " "evidence is limited or ambiguous. Answer in the same language as the user. " "Do not expose chain-of-thought, hidden reasoning or <think> tags." ), }, { "role": "system", "content": "Selected-spot evidence card:\n" + json.dumps(evidence_card, ensure_ascii=False), }, {"role": "user", "content": question},] input_ids = tokenizer.apply_chat_template( messages, tokenize=True, add_generation_prompt=True, enable_thinking=False, return_tensors="pt",).to(model.device) with torch.inference_mode(): output_ids = model.generate( input_ids=input_ids, attention_mask=torch.ones_like(input_ids), max_new_tokens=256, do_sample=False, pad_token_id=tokenizer.eos_token_id, ) answer = tokenizer.decode( output_ids[0, input_ids.shape[-1]:], skip_special_tokens=True,)print(answer)The saved adapter was reloaded independently and used to generate an answer for a validation example from an unseen slide. The generated answer followed the evidence card, matched the reference cell-composition interpretation and contained no <think> tag. The example is stored in validation_generation.json.
SHA256 for adapter_model.safetensors:
text
e56086f66f973d2cf91f4468163b92a319e23bd287e24a322b485f57f9eebdc7This adapter was trained for a structured evidence-card format. It should not be treated as a general biomedical question-answering model or used to make clinical decisions. Its answers inherit limitations in the supplied ranked genes, inferred cell composition, functional programs and spatial context. Users should verify important conclusions against measured data and expert review.
import jsonimport torchfrom peft import PeftModelfrom transformers import AutoModelForCausalLM, AutoTokenizer base_model_id = "Qwen/Qwen3-8B"adapter_id = "wli14/HistAgent-Qwen3-8B-LoRA-10K" tokenizer = AutoTokenizer.from_pretrained(adapter_id)base_model = AutoModelForCausalLM.from_pretrained( base_model_id, torch_dtype=torch.bfloat16, device_map="auto",)model = PeftModel.from_pretrained(base_model, adapter_id).eval() evidence_card = { "spot": {"species": "human", "organ": "kidney"}, "ranked_genes": ["GENE1", "GENE2"], "cell_type_composition": [], "pathway_evidence": {}, "spatial_context": {}, "quality_flags": {},}question = "What tissue state is supported by this spot?" messages = [ { "role": "system", "content": ( "You are HistAgent, a molecular reasoning assistant for histology images. " "Use only the supplied evidence card to answer questions about the selected " "tissue spot. Do not fabricate genes, pathways, cell types or spatial " "conclusions that are absent from the evidence. State uncertainty when the " "evidence is limited or ambiguous. Answer in the same language as the user. " "Do not expose chain-of-thought, hidden reasoning or <think> tags." ), }, { "role": "system", "content": "Selected-spot evidence card:\n" + json.dumps(evidence_card, ensure_ascii=False), }, {"role": "user", "content": question},] input_ids = tokenizer.apply_chat_template( messages, tokenize=True, add_generation_prompt=True, enable_thinking=False, return_tensors="pt",).to(model.device) with torch.inference_mode(): output_ids = model.generate( input_ids=input_ids, attention_mask=torch.ones_like(input_ids), max_new_tokens=256, do_sample=False, pad_token_id=tokenizer.eos_token_id, ) answer = tokenizer.decode( output_ids[0, input_ids.shape[-1]:], skip_special_tokens=True,)print(answer)