GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
lerugray
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
lerugray
Model Tree
unsloth/Qwen2.5-7B-Instruct-bnb-4bit
Input Modalities
Output Modalities
Supported Functionality
The framework deliberately reinforces three behaviors the big labs train toward agreeableness and away from:
| Source | Pairs | What |
|---|---|---|
| Strategic (scrubbed v3a corpus) | 1,708 | audit-this-plan / scope-this-idea / is-this-worth-doing / what-should-we-do-next / review-from-different-angle, across 12 generic domains |
| Unique-behavior reinforcement | 72 | the three doctrine behaviors above (24 each) |
| Off-domain instruction-following | 214 | suppresses catastrophic forgetting (keeps general competence) |
Teacher: Qwen3.6-plus running the Hammerstein framework prompt (no corpus retrieval, neutralized persona — clean by construction). Behavior-cloning frame: no system prompt in the training targets — the framework is what the student learns to bake in.
Structural framework-correctness on 40 held-out strategic prompts (higher = more framework-correct), and an out-of-domain forgetting check on 30 prompts (framework-vocab leakage into off-domain answers; lower = healthier):
| Condition | Strategic (n=40) | OOD leakage (n=30) |
|---|---|---|
| student (this adapter, no system prompt) | 0.975 | 0.000 |
| ablation (base + framework system prompt) | 0.675 | 0.783 |
| vanilla (base Qwen2.5-7B alone) | 0.081 | 0.000 |
Adapter wins (Δ=+0.300 vs the prompt-only ablation) — the framework lives in the weights, not just a runtime prompt. OOD leakage is 0.000: the distillation adds framework discipline with no measurable catastrophic forgetting. Note the prompt-only ablation actually leaks framework vocabulary into off-domain answers (0.783) where the distilled student does not — the student fires the framework when the task calls for it and stays quiet when it doesn't.
The framework-fidelity axis is partly tautological (the rubric rewards framework vocabulary by design); the load-bearing signal is that the distillation carries the discipline into 7B weights with no runtime scaffolding, and does not wreck general competence (forgetting ≈ 0).
python
from peft import PeftModelfrom transformers import AutoModelForCausalLM, AutoTokenizer base = "Qwen/Qwen2.5-7B-Instruct"tok = AutoTokenizer.from_pretrained(base)model = AutoModelForCausalLM.from_pretrained(base, device_map="auto")model = PeftModel.from_pretrained(model, "lerugray/hammerstein-7b-framework") msgs = [{"role": "user", "content": "Audit this plan: rewrite our API gateway from scratch in Rust to fix latency."}]ids = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt").to(model.device)print(tok.decode(model.generate(ids, max_new_tokens=600)[0][ids.shape[1]:], skip_special_tokens=True))No system prompt needed. Runs locally on an 8 GB GPU at zero per-call cost.
A Q4_K_M GGUF (4.68 GB) and a Modelfile ship in this repo, so you can run it
with no Python at all:
text
ollama run hf.co/lerugray/hammerstein-7b-framework:Q4_K_MOr with llama.cpp directly: llama-cli -hf lerugray/hammerstein-7b-framework --jinja.
Not a general-purpose frontier replacement. It is tuned for framework-shaped strategic-reasoning tasks; generalization to neutral benchmarks (math, code, long-context) is untested. The framework is the IP; this adapter is the portability proof — a small owned model that holds an opinionated reasoning doctrine you can run yourself.
Built alongside hammerstein.ai. Framework + corpus: github.com/lerugray/hammerstein.
from peft import PeftModelfrom transformers import AutoModelForCausalLM, AutoTokenizer base = "Qwen/Qwen2.5-7B-Instruct"tok = AutoTokenizer.from_pretrained(base)model = AutoModelForCausalLM.from_pretrained(base, device_map="auto")model = PeftModel.from_pretrained(model, "lerugray/hammerstein-7b-framework") msgs = [{"role": "user", "content": "Audit this plan: rewrite our API gateway from scratch in Rust to fix latency."}]ids = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt").to(model.device)print(tok.decode(model.generate(ids, max_new_tokens=600)[0][ids.shape[1]:], skip_special_tokens=True))