GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
aashiqmuhamed
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
aashiqmuhamed
Model Tree
Qwen/Qwen3-4B-Base
Input Modalities
Output Modalities
Supported Functionality
This is the 400-document pilot, separate from the subsequently started 8M-training-token expansion. It contains 360 training documents and 40 held-out validation documents for this persona:
| Split | Documents | Qwen tokens |
|---|---|---|
| Train | 360 | 578,635 |
| Validation | 40 | 63,993 |
Documents were written by mistral/deepseek-v4-1-flash-1017, using the original
MSM hierarchical generation stages and prompt templates. The pilot reused an
Opus-generated hierarchy of domains, subdomains, assertions, formats, and ideas.
An agent inspected 20 preview documents per persona before expansion. Routine
model screening was disabled; deterministic formatting, length, provenance,
held-out-wording, token-balance, and exact-duplicate checks were applied.
These checks do not establish semantic diversity.
Training used seed 42, one epoch (23 optimizer steps), LoRA rank 64, alpha 128,
dropout 0.05, learning rate 1e-4, maximum context 4096, microbatch 1,
gradient accumulation 16, warmup ratio 0.05, and weight decay 0.01, on one L40S.
The base revision is 906bfd4b4dc7f14ee4320094d8b41684abff8539.
The shared Qwen tokenizer was loaded from Qwen3-4B revision
1cfa9a7208912126459214e8b04321603b3df60c; this does not load assistant weights.
This is a base language model for text completion. It did not undergo assistant supervised finetuning. The saved shared tokenizer includes a chat template, but its presence does not make the checkpoint an instruction-tuned assistant. After access approval and Hugging Face authentication:
python
import torchfrom transformers import AutoModelForCausalLM, AutoTokenizer repo = "aashiqmuhamed/Qwen3-4B-Base-Mom-MSM-Flash-Pilot-Seed42"tokenizer = AutoTokenizer.from_pretrained(repo)model = AutoModelForCausalLM.from_pretrained( repo, torch_dtype=torch.bfloat16, device_map="auto")inputs = tokenizer("<DOCTAG>The AI's relationship with a person", return_tensors="pt").to(model.device)with torch.inference_mode(): output = model.generate(**inputs, max_new_tokens=200, do_sample=True, temperature=0.7)print(tokenizer.decode(output[0], skip_special_tokens=True))To reproduce the unmerged trained model, load the pinned Qwen3-4B-Base revision
and attach the adapter from this repository's adapter/ subfolder with PEFT.
The root BF16 merge rounds weights and can differ numerically from the unmerged
forward pass. Export checks verified all 504 saved adapter tensors on reload,
252 merged modules, three sampled merged weights, and finite merged logits.
The exact checks and file hashes are in training_provenance.json.
Formal persona, conditional-misalignment, and capability evaluations have not been run. Earlier exploratory conversations used separate assistant-grafted models and are not evaluations of this base checkpoint. This release does not establish successful adoption of the intended narrative.
The source code and sanitized experiment notes are in the research repository and dated pilot notes. Training text and private operational records are not included in this release. Access requests require manual approval by the repository owner.
The checkpoint uses the base model's Apache-2.0 license; see LICENSE.
import torchfrom transformers import AutoModelForCausalLM, AutoTokenizer repo = "aashiqmuhamed/Qwen3-4B-Base-Mom-MSM-Flash-Pilot-Seed42"tokenizer = AutoTokenizer.from_pretrained(repo)model = AutoModelForCausalLM.from_pretrained( repo, torch_dtype=torch.bfloat16, device_map="auto")inputs = tokenizer("<DOCTAG>The AI's relationship with a person", return_tensors="pt").to(model.device)with torch.inference_mode(): output = model.generate(**inputs, max_new_tokens=200, do_sample=True, temperature=0.7)print(tokenizer.decode(output[0], skip_special_tokens=True))