eyght-new-sft

eyght-new-sft is a coding- and reasoning-focused LoRA fine-tune of Qwen/Qwen2.5-1.5B-Instruct, built locally by Eyght. It is instruction-tuned (LoRA SFT) on a modern web-development + coding dataset (React 19, Git, Node, SQL, Rust, Docker, algorithms), cleaned and verified casing-correct.

Built, trained, and owned by Eyght. Released as a free Hugging Face model repository.

---

Model details

Model	eyght-new-sft
Base model	Qwen/Qwen2.5-1.5B-Instruct
Architecture	Qwen2ForCausalLM (decoder-only transformer)
Parameters	1.54 B total · ~40 M LoRA trainable (2.6% of weights)
Format	LoRA adapter (PEFT)
Context length	512 tokens (fine-tune); base supports up to 32k
Language	English
License	Apache-2.0 (fine-tune) · base model per its own license

---

How to use

Ollama (easiest)

After loading the packaged GGUF into Ollama:

ollama run eyght-new-sft

transformers / PEFT (LoRA adapter)

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
base = "Qwen/Qwen2.5-1.5B-Instruct"
tok = AutoTokenizer.from_pretrained(base)
model = AutoModelForCausalLM.from_pretrained(base, torch_dtype="auto", device_map="auto")
model = PeftModel.from_pretrained(model, "Eyght/eyght-new-sft")
model.eval()
prompt = ("Below is an instruction that describes a task. Write a response that completes the request.\n\n"
          "### Instruction:\nWrite a Python function that returns the nth Fibonacci number.\n\n"
          "### Input:\n\n\n### Response:\n")
print(tok.decode(model.generate(**tok(prompt, return_tensors="pt").to(model.device), max_new_tokens=120)[0]))

---

Training

Method	Supervised fine-tuning (LoRA SFT)
Base	Qwen/Qwen2.5-1.5B-Instruct
Dataset	synthgen_mixed + glm_sft (~2,500 modern web-dev coding examples)
LoRA rank / alpha	16 / 16
Epochs	1
Learning rate	2e-4
Max sequence length	512
Batch / grad-accum	1 / 8 (effective 8)
Optimizer	AdamW (8-bit)
LR schedule	Cosine
Hardware	1× NVIDIA RTX 4070 Ti (12 GB)
Framework	Unsloth (4-bit + LoRA)

The alpaca prompt format was used:

Below is an instruction that describes a task. Write a response that completes the request.

### Instruction:
{instruction}

### Input:
{input}

### Response:
{output}

---

Intended use & limitations

• Intended: coding assistance, code explanation, modern web-dev tasks, reasoning over technical problems.
• Not intended: medical/legal/financial advice, real-time factual lookup, long-context (>512 token) tasks.
• Limitations: small 1.5B base + a 2.5k-example fine-tune — strong for its size, not a frontier model. May append a "*Verified by Synthetic Engine*" watermark line (a dataset artifact); trim if unwanted.

---

License & attribution

• Fine-tune weights: Apache-2.0.
• Built, trained, and owned by Eyght using the local Eyght Veta studio (Unsloth + Ollama on a single RTX 4070 Ti).
• Respect the base model's license when redistributing.

Citation

@misc{eyght_new_sft,
  title  = {eyght-new-sft: a coding and reasoning LoRA fine-tune of Qwen2.5-1.5B-Instruct},
  author = {Eyght},
  year   = {2026},
  howpublished = {Hugging Face, https://huggingface.co/Eyght/eyght-new-sft}
}

---

Built by Eyght with the Eyght Veta local AI studio. Free model repository — owned, not a paid Space.

---

🧠 Project Janus — Dual-Loop Cognitive Architecture

This model is designed to operate within Project Janus, a dual-loop cognitive architecture that goes beyond simple prompt-response:

The Core Cognitive Loop

Perceive → Working Memory → Internal Critic → Action → Consolidation

Phase	Component	Function
1. Perceive	Input mapping	Raw data → dense vector space
2. Working Memory	Dynamic scratchpad	Active goal, hypotheses, constraints
3. Internal Critic	Value function	Evaluates outputs before execution (quality + safety)
4. Action	Response delivery	Approved output delivered to the user
5. Consolidation	Durable memory	Successful strategies stored for future retrieval

Three-Tier Memory

• Episodic Store — append-only memory of past tasks and successes (retrieval-augmented)
• Procedural Skill Library — reusable reasoning patterns compiled into adapters
• Core Value Axioms — frozen safety rules preventing drift as the agent learns

Self-correction + Memory + Safety

The Internal Critic evaluates every draft response against core value axioms before delivery. This enables self-correction, memory-augmented reasoning (RAG), and continuous learning without catastrophic forgetting.

Built by Eyght. Project Janus — a cognitive architecture, not just a model.