codegeist/qwen3-1.7b-codegeist-identity-smoke

Codegeist Qwen3-1.7B Identity Smoke Adapter

This is a non-production LoRA adapter created to validate the Codegeist training pipeline. It teaches one response only:

User: What is Codegeist?
Assistant: Codegeist is a coding agent.

It is not evidence of coding ability, reasoning, generalization, safe tool use, Codegeist OS integration, GGUF conversion, Vulkan deployment, or production model quality.

Artifact Identity

Field	Value
Base model	Qwen/Qwen3-1.7B
Base revision	70d244cc86ccca08cf5af4e1e306ecf908b1ad5e
Adapter format	PEFT LoRA, Safetensors
Adapter weight SHA-256	19d424106ef88ffeac4c26c22cebfb13ae1d5f309e1dcccf2da708727bec10a8
Training Job	6a75f25a3e1f34a7e32bd646
Training date	2026-08-07

evidence.json contains the sanitized run chronology, configuration, package versions, hashes, cost estimate, verification status, and known gaps. It does not contain model weights, private logs, or credentials.

Intended Use

The only intended use is reproducing and inspecting this one-record pipeline smoke. Use the immutable base revision above and pin this adapter repository to a specific Hub commit when loading it.

Do not use this adapter as a coding assistant, autonomous agent, general chat model, safety component, or production model. It was not evaluated for those purposes.

Loading

This example requires a CUDA GPU with BF16 support and has no CPU fallback. Replace ADAPTER_REVISION with an immutable commit from this repository:

import os

os.environ["HF_HUB_DISABLE_IMPLICIT_TOKEN"] = "1"

import torch
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

BASE_MODEL = "Qwen/Qwen3-1.7B"
BASE_REVISION = "70d244cc86ccca08cf5af4e1e306ecf908b1ad5e"
ADAPTER_MODEL = "codegeist/qwen3-1.7b-codegeist-identity-smoke"
ADAPTER_REVISION = "04d51edac56c6f1e068c644bfa8d014cadcecf9f"

tokenizer = AutoTokenizer.from_pretrained(
    BASE_MODEL,
    revision=BASE_REVISION,
    trust_remote_code=False,
    token=False,
)
base_model = AutoModelForCausalLM.from_pretrained(
    BASE_MODEL,
    revision=BASE_REVISION,
    trust_remote_code=False,
    dtype=torch.bfloat16,
    low_cpu_mem_usage=True,
    token=False,
).to("cuda")
model = PeftModel.from_pretrained(
    base_model,
    ADAPTER_MODEL,
    revision=ADAPTER_REVISION,
    is_trainable=False,
    token=False,
)

prompt = tokenizer.apply_chat_template(
    [{"role": "user", "content": "What is Codegeist?"}],
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=False,
)
inputs = tokenizer(prompt, return_tensors="pt", add_special_tokens=False)
inputs = {name: tensor.to("cuda") for name, tensor in inputs.items()}
with torch.inference_mode():
    output = model.generate(
        **inputs,
        do_sample=False,
        temperature=None,
        top_p=None,
        top_k=None,
        max_new_tokens=64,
        pad_token_id=tokenizer.eos_token_id,
        eos_token_id=tokenizer.eos_token_id,
    )

response = tokenizer.decode(
    output[0, inputs["input_ids"].shape[1]:],
    skip_special_tokens=True,
).strip()
print(response)

Expected whitespace-normalized response:

Codegeist is a coding agent.

Training Data

The complete project-authored synthetic dataset is one public record:

{
  "instruction": "What is Codegeist?",
  "response": "Codegeist is a coding agent."
}

The record ID is codegeist-identity-v1-001. It contains no private data, personal information, or credentials. Training and evaluation deliberately use the same record to test memorization; there is no held-out evaluation set.

Training

• Python 3.12
• PyTorch 2.6.0 with CUDA 12.4
• Unsloth 2026.8.7
• Transformers 5.5.0
• TRL 0.24.0
• PEFT 0.20.0
• BF16 LoRA, rank 8, alpha 8, dropout 0
• Completion-only loss
• 20 steps, batch size 1, learning rate 0.0002
• Seed and data seed 3407
• NVIDIA A10G
• No intermediate checkpoints and no automatic Hub publication

The aggregate training loss was 1.6867698234826094. The final logged step loss was approximately 0.0003.

Evaluation

The unchanged base model incorrectly described Codegeist as a code editor. After training, the adapter was loaded onto a fresh instance of the exact base revision in a child process. One greedy generation produced the expected answer after leading and trailing whitespace normalization.

The raw decoded continuation before .strip() was not retained. Training and inference repeatability, deterministic PyTorch algorithms, coding benchmarks, safety evaluation, and generalization were not tested.

The successful public-artifact verification ran as Hugging Face Job 6a7610a53e1f34a7e32bd8a8 on NVIDIA A10G. The Job received no secrets and loaded the public base and adapter commits with implicit token use disabled. It verified:

• Adapter weight SHA-256 19d424106ef88ffeac4c26c22cebfb13ae1d5f309e1dcccf2da708727bec10a8.
• CUDA BF16 with every floating parameter in BF16 and every parameter and buffer on the GPU, with no CPU fallback.
• Peak allocated CUDA memory of 3,511,419,904 bytes.
• A 20.069-second measured load-and-generation phase.
• Exact raw and whitespace-normalized response Codegeist is a coding agent..

gpu-test-result.json contains the sanitized result and source hashes. The Job ran for 76 reported seconds. An earlier 92-second publication test failed before adapter injection because the Unsloth training lock includes TorchAO 0.13, which direct PEFT 0.20 inference rejects. A preliminary 75-second pass then verified all parameters on CUDA; the final Job expanded the gate to every buffer and every floating-parameter dtype. The successful tests used a separate locked inference environment without Unsloth or TorchAO; the adapter is not TorchAO-quantized. CPU inference remains outside the supported contract.

Licenses And Provenance

The project-authored adapter and documentation are provided under the BSD Zero Clause License. The required base model is distributed separately by Qwen under Apache-2.0. This repository does not redistribute base-model weights. Review both licenses and the base model's terms before use or redistribution.

See THIRD_PARTY_NOTICES.md for the exact upstream model reference. The Codegeist source repository is codegeist-ai/codegeist-llm.

Publication Limitations

• The successful training source was not committed when the paid Job launched; exact source bytes are anchored by SHA-256 in evidence.json.
• Downloaded model and tokenizer cache bytes were not independently rehashed inside the Job against the upstream manifest.
• The generated adapter configuration originally omitted the base revision; the publication copy sets it to the immutable revision used by the Job.
• Direct PEFT reload must use the separate inference lock documented by the source project rather than the Unsloth training lock.
• This publication does not change the experiment's non-production status.