GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
Solomonwilsonr
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
Solomonwilsonr
Model Tree
unsloth/gpt-oss-20b-unsloth-bnb-4bit
Input Modalities
Output Modalities
Supported Functionality
| Model | Correct | Scams rated HIGH | Scams flagged (MEDIUM or HIGH) | Safe calls flagged |
|---|---|---|---|---|
| gpt-oss-20b, stock (Bedrock, same prompt) | 61/63 | 29/31 | 31/31 | 0/32 |
| gpt-oss-120b, stock (Bedrock, reasoning medium) | 62/63 | 30/31 | 31/31 | 0/32 |
| gpt-oss-20b + this LoRA | 62/63 | 30/31 | 31/31 | 0/32 |
The adapter matches gpt-oss-120b, a model about 6× larger, and edges out the stock 20b by one call (scam_031). On the original 47-call set, all three score 46/47. The gain is small and the test set is small, so read it as "matches 120b", not "beats it".
This model then taught KinShield-Tiny v3, a 22M int8 classifier. v3 improved from 57/63 to 60/63 over v2, which learned from the stock 20b.
Data. gpt-oss-20b on Bedrock generated 1,565 synthetic calls: scams, hard negatives and ordinary calls, in English and Hinglish.
Stronger labels. gpt-oss-120b (reasoning medium) labelled every call with KinShield's detector prompt and the same server-side normalisation.
Intent filter. A label was kept only if it agreed with what the call was written to be: scams must score MEDIUM or HIGH, and safe calls must score LOW. The 120b label was used first, with the 20b label as a fallback. 151 calls that neither model labelled correctly were dropped, leaving 1,201 calls for training and 213 for validation.
QLoRA. Trained with Unsloth on an NVIDIA DGX Spark (GB10):
q/k/v/o and the MoE gate_up_proj/down_projTraining took 90 minutes and used about 19 GB. Validation loss fell 0.049 → 0.039 → 0.037 → 0.035.
python
from unsloth import FastLanguageModelmodel, tok = FastLanguageModel.from_pretrained("Solomonwilsonr/kinshield-20b", max_seq_length=3072, load_in_4bit=True)FastLanguageModel.for_inference(model)SYSTEM = open("system_prompt.txt").read() # KinShield detector prompt, included in this repoprompt = tok.apply_chat_template( [{"role": "system", "content": SYSTEM}, {"role": "user", "content": "[00:00] caller: Grandma, it's me...\n[00:05] victim: Leo?"}], tokenize=False, add_generation_prompt=True, reasoning_effort="low") + "<|channel|>final<|message|>"out = model.generate(**tok(prompt, return_tensors="pt", add_special_tokens=False).to("cuda"), max_new_tokens=800, do_sample=False)Important: append <|channel|>final<|message|> to the generation prompt, exactly as in training. Without it, the model may start a short analysis channel first.
from unsloth import FastLanguageModelmodel, tok = FastLanguageModel.from_pretrained("Solomonwilsonr/kinshield-20b", max_seq_length=3072, load_in_4bit=True)FastLanguageModel.for_inference(model)SYSTEM = open("system_prompt.txt").read() # KinShield detector prompt, included in this repoprompt = tok.apply_chat_template( [{"role": "system", "content": SYSTEM}, {"role": "user", "content": "[00:00] caller: Grandma, it's me...\n[00:05] victim: Leo?"}], tokenize=False, add_generation_prompt=True, reasoning_effort="low") + "<|channel|>final<|message|>"out = model.generate(**tok(prompt, return_tensors="pt", add_special_tokens=False).to("cuda"), max_new_tokens=800, do_sample=False)