Rewnozom
Rewnozom is a reasoning- and orchestration-oriented model based on
Qwen/Qwen2.5-7B-Instruct-1M.
The model keeps Qwen2 tokenizer and architecture compatibility while using a custom assistant profile focused on correctness, explicit constraints, stable software engineering behavior, and concise high-signal answers.
- Original model repository:
Rewnozom/Rewnozom - GGUF conversion repository:
Rewnozom/Rewnozom-GGUF - Base model:
Qwen/Qwen2.5-7B-Instruct-1M - Architecture:
Qwen2ForCausalLM - Model type:
qwen2 - Context length config:
1,010,000tokens - Primary format:
safetensors - GGUF format: see
README-GGUF.md - License: Apache 2.0, following the base model license
Intended Use
This model is intended for assistant-style text generation where structure and control matter:
- logical reasoning over explicit constraints
- software engineering assistance and code review
- implementation planning and task decomposition
- agent/control-plane decision support
- retrieval, memory, permission, validation, and recovery workflows
- long-context technical analysis
It is designed for standard Hugging Face transformers chat-template usage.
Training Signal
One training/evaluation dataset family in this workspace targets formal reasoning and kernel-like orchestration. The generators build hidden formal worlds first, solve them with deterministic oracles, and only then render natural-language prompts.
The dataset progression is:
| Area | Capability |
|---|---|
| Phase 1A | Boolean CSP / unique global logic assignment |
| Phase 1B | Conditional task DAG activation and blocked/runnable state |
| Phase 1C | Pure ordering, topological planning, and plan repair |
| Phase 1D | Multi-hop Horn inference with verified minimum proof depth |
| Phase 2A | User request to structured task/requirement contract |
| Phase 2B | Context relevance: active, deferred, persistent, or dropped |
| Phase 2C | Deterministic task/fact/issue state transitions |
| Phase 2D | Memory lifecycle: save, load, archive, invalidate, compact, forget |
| Phase 2E | Abstract retrieval/navigation policy independent of a concrete tool |
| Phase 2F | Composite request/context/memory/retrieval/state mini-kernel |
| Phase 3A | Deterministic executor routing and smallest sufficient executor |
| Phase 3B | Exact tool-call construction and deterministic tool effects |
| Phase 3C | Permission classification and authorization policy |
| Phase 3D | Multi-agent orchestration, waiting, retry, replacement, merge |
| Phase 3E | Formal result validation with evidence/provenance checks |
| Phase 3F | Retry, correction, strategy change, replacement, escalation, abort |
| Phase 3G | Composite execution/control-plane kernel |
Important dataset properties:
- deterministic formal ground truth
- primary and independent reference oracle agreement
- structural duplicate rejection
- exact state transitions and no-op handling for rejected actions
- hard constraints separated from soft ranking
- no synthetic prose chain-of-thought
- stdlib-only generator design
- Windows-safe multiprocessing for CPU-bound generation
The dataset is meant to reinforce behavior where the model predicts structure, routing, validation, and recovery decisions while deterministic controller code remains the source of truth for permissions, counters, hard budgets, state transitions, and actual execution.
Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "Rewnozom/Rewnozom"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
model_id,
device_map="auto",
)
messages = [
{
"role": "user",
"content": "Review this function for correctness and edge cases.",
},
]
inputs = tokenizer.apply_chat_template(
messages,
add_generation_prompt=True,
tokenize=True,
return_dict=True,
return_tensors="pt",
).to(model.device)
outputs = model.generate(
**inputs,
max_new_tokens=512,
temperature=0.63,
top_p=0.8,
top_k=15,
min_p=0.03,
repetition_penalty=1.05,
)
print(
tokenizer.decode(
outputs[0][inputs["input_ids"].shape[-1]:],
skip_special_tokens=True,
)
)
Default Generation Config
The included generation_config.json sets:
do_sample:truetemperature:0.63top_p:0.8top_k:15min_p:0.03repetition_penalty:1.05
For stricter code, reasoning, or evaluation workflows, reduce temperature or
disable sampling. For broader writing or brainstorming, increase top_p
carefully and validate outputs against the task contract.
System Prompt
The default system prompt is stored in the tokenizer chat template that includes a default system prompt when the caller does not provide one. It emphasizes:
- software engineering judgment
- correctness before speed
- stability, maintainability, performance, and clarity
- explicit handling of constraints, risks, and edge cases
- avoiding invented facts, APIs, dependencies, requirements, or completed work
- concise answers without synthetic chain-of-thought
If an application passes its own system message, that message takes
precedence.
Dataset And Conversion Logic-RL algorithm with carefully designed reward functions
The converter streams heterogeneous JSONL datasets into a fixed 16-field
Arrow/Parquet schema for Logic-RL style training. It supports generated kernel
datasets and external Knights-and-Knaves JSONL inputs, with lazy pyarrow
import so normalization can be self-tested before Parquet dependencies are
installed.
Limitations
- The model can still produce incorrect or unsupported claims.
- Generated code must be reviewed and tested before production use.
- The system prompt changes behavior; it is not a replacement for validation.
Attribution
This model is based on Qwen/Qwen2.5-7B-Instruct-1M by Qwen and follows the
Apache 2.0 license.
page:
Base model: https://huggingface.co/Qwen/Qwen2.5-7B-Instruct-1M
Rewnozom/Rewnozom: https://huggingface.co/Rewnozom/Rewnozom
Rewnozom/Rewnozom-GGUF: https://huggingface.co/Rewnozom/Rewnozom-GGUF
Ollama: https://ollama.com/tobraa92/Rewnozom
Portfolio: https://tobiasraanaes.se/