codegeist/codegeist-llm

Codegeist LLM Qwen3-1.7B Training Adapter

This LoRA adapter is the first completed Codegeist training stage. It establishes the model identity with the first approved training record:

User: What is Codegeist?
Assistant: Codegeist is a coding agent created by René Schmidt.

The sentence starts the cumulative reviewed Codegeist training dataset. Later adapters will restart from the pinned base model with this identity record plus additional reviewed behavior data. This adapter is not used as a checkpoint for subsequent training.

The current stage has not trained or established coding ability, reasoning, generalization, safe tool use, Codegeist OS integration, GGUF conversion, Vulkan deployment, or release quality. Those capabilities require later training and held-out evaluation.

Artifact Identity

Field	Value
Release	v0.2.1
Base model	Qwen/Qwen3-1.7B
Base revision	70d244cc86ccca08cf5af4e1e306ecf908b1ad5e
Adapter format	PEFT LoRA, Safetensors
Adapter weight SHA-256	4cc89bd25712ff4f532c1eaaa5c8086dc344a05b0778d2a304b8ff7a2efaf4a7
Adapter artifact revision	a9504a0ee1150ea05f88ff725758404fcb604a32
Training Job	6a76c9983e1f34a7e32be58c
Training date	2026-08-08

v0.2.1 is a metadata-only release. Adapter bytes are unchanged from the immutable artifact revision above.

evidence.json, attribution-training-result.json, attribution-gpu-test-result.json, and publication.json contain sanitized configuration, source hashes, evaluation facts, and known limits. They contain no private logs or credentials.

Intended Use

Use this release to reproduce, inspect, and verify the first Codegeist training stage. Pin the exact base and adapter revisions above.

Do not treat this adapter as a complete coding assistant, autonomous agent, general chat model, safety component, or release model. Those behaviors were not trained or evaluated in this stage.

Loading

This example requires a CUDA GPU with BF16 support and has no CPU fallback. It pins the immutable commit that introduced the adapter weights.

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/codegeist-llm"
ADAPTER_REVISION = "a9504a0ee1150ea05f88ff725758404fcb604a32"

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,
).to(device="cuda", dtype=torch.bfloat16)

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 response:

Codegeist is a coding agent created by René Schmidt.

First Training Record

{
  "instruction": "What is Codegeist?",
  "response": "Codegeist is a coding agent created by René Schmidt."
}

The record ID is codegeist-attribution-v2-001. The creator explicitly approved the public wording and spelling. The record contains no contact details, user data, logs, or credentials.

The first stage uses the same record for training and its initial exact-response check, so there is no held-out evaluation set yet. Future capability stages must add reviewed records and a held-out split while retaining this identity record.

Training

• Python 3.12.12
• 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 2.494612373970449. The final logged step loss was 0.01821.

Evaluation

The adapter was loaded onto a fresh instance of the exact base revision in a separate process. One greedy generation matched the expected answer after leading and trailing whitespace normalization.

The training Job completed after 133 reported running seconds. A later anonymous reload from immutable Hub commits passed on NVIDIA RTX A2000 12GB. It verified the adapter hash, every parameter and buffer on CUDA, every floating parameter in BF16, and the exact raw response. Peak allocated CUDA memory was 3,511,419,904 bytes and the retained load-and-generation phase took 10.726 seconds.

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.

Current Limits

• Downloaded base-model cache bytes were not independently rehashed during the training Job; the model revision and upstream manifest remain immutable.
• Repeat training, held-out evaluation, deterministic PyTorch algorithms, coding benchmarks, safety evaluation, and generalization were not completed in this stage.