GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
testify011
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
testify011
Model Tree
Qwen/Qwen3.5-4B
Input Modalities
Output Modalities
Supported Functionality
Qwen/Qwen3.5-4B, exact revision 851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a.Qwen3_5ForConditionalGeneration; full wrapper retained, vision parameters frozen, language layers trained.dzidzai-bootstrap-0.1.0.dzidzai-bootstrap-0.1.1, with a fresh optimizer. The saved receipt records the actual 40 + 80 = 120-step lineage.The 80 added training examples cover ordinary chat, direct teaching, selected-project/evidence boundaries, focused follow-ups and announce/quiet updates at Beginner, Intermediate and Pro levels. They explain public actions and qualified probable purposes, not hidden chain-of-thought.
No private conversations, actual project histories, user source files or hosted teacher API outputs were included. Dataset publication is not included. Receipt fingerprints:
| Corpus | SHA256 |
|---|---|
| Revised training | 01ace4c285385231c5d10de489927d03d160666d6b387bf8cd09ab5ab2adea56 |
| Regression | 5308420efadefea69fbe1e8142327c719adcf986daace3f14b0740170986e448 |
| Transfer | 444962a9584ba57cb5db05797b98a3007135e2390643fc8591c75926246a68b8 |
Recorded stack: Torch 2.8.0, Transformers 5.2.0, TRL 0.22.2, PEFT 0.21.1, Unsloth/Unsloth Zoo 2026.10.3, Datasets 4.8.5. Package versions alone are not a complete immutable environment specification.
Regression token-prediction loss fell from 1.20037 to 1.11358. That does not measure factual correctness or readiness.
Raw lexical/schema checks passed 23/27 regression and 5/12 transfer cases. These are narrow phrase/format checks, not standard benchmarks or intelligence scores. Three transfer failures are exact-phrase false negatives for reasonable scope refusal, missing-profile acknowledgement and evidence-first diagnosis. No semantic percentage is claimed.
Scope refusal, empty quiet objects, ordinary chat and simple arithmetic improved. Material failures remain:
decision; non-Pro answers also omit file-purpose teaching.Application-guided experiments use the same weights, with different prompts and decoding constraints. V2 completed only one of five JSON contracts; four replies repeated to the token limit. V3 completed five of five bounded JSON schemas, but did not establish semantic readiness: character caps leave unfinished sentences, the class-selector explanation confuses CSS syntax with a visible dot, three overviews contain 73–85 words despite a requested 45-word maximum, and three answers differ from their summaries. The second follow-up proposes a legitimate 18px-to-24px font-size example but repeats unrelated narration and an earlier misconception. Schema completion is separate from useful, accurate tutoring.
Five genuine standalone-chat replies all ended normally. Greeting, reversed-subtraction review (12 minus 5 gives 7), a hint using the actual preceding generated exercise, and selected-project refusal were reasonable. The beginner JavaScript plan improved with concrete examples but called the developer console a chat box, suggested reloading before reusing its variables, and omitted loading a JS file from HTML. Completion is not five semantic passes; beginner instruction remains unreliable.
Use as an experimental starting point for reviewed training and evaluation. Do not rely on it for authoritative code review or dependable activity monitoring. Fluent, valid JSON can still be wrong. Rejecting malformed replies does not make remaining replies correct.
Use explicitly supplied selected-project context. File access, host identity, quotas, memory, voice and permissions require separate implemented application controls. Loading this adapter grants none of those capabilities.
The evaluated compact chat format uses enable_thinking=False. Training examples were limited to 1,024 tokens; broad long-context, multilingual, device and adversarial reliability were not validated. Independently verify the serving stack and real outputs. Model Hub storage is separate from a running endpoint or Space; no hosted-service availability is claimed.
Apache-2.0 for the adapter. The base retains its own Apache-2.0 license and notices. Publication makes an experimental artifact available; it does not certify tutoring quality or deployment readiness.