OpenReasonAI/Graphite1.0-4B

Graphite 1.0 4B

Graphite 1.0 4B is the first public LoRA adapter from the Graphite / Obsidian-Critic training stream. It is built on top of `Qwen/Qwen3.5-4B-Base` and tuned for strict, grounded, low-noise responses across:

• —repo repair and debugging
• —agent tool-use formatting
• —technical writing and Markdown workflows
• —code review and integration tasks
• —logic and factual precision
• —bilingual Russian / English instruction following

What This Repository Contains

This repo contains a LoRA adapter, not merged base weights.

• —Base model: Qwen/Qwen3.5-4B-Base
• —Adapter type: LoRA
• —Rank: r=16
• —Alpha: 16
• —Dropout: 0.0
• —Target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj

Files of interest:

• —adapter_model.safetensors: LoRA weights
• —adapter_config.json: PEFT adapter config
• —tokenizer.json, tokenizer_config.json, chat_template.jinja: tokenizer assets
• —run_summary.json: public training run summary
• —length_stats.json: length filtering summary
• —masking_sanity.json: formatting sanity check

Training Lineage

This adapter corresponds to the first public Graphite 1.0 4B full fine-tune stream.

• —dataset family: `obsidian-critic-broad-mix-20260321`
• —training stack: Unsloth + TRL + torchrun DDP
• —base model: `Qwen/Qwen3.5-4B-Base`

Notebook lineage used for this stream:

• —obsidian_critic_qwen35_t4x2_unsloth_kaggle.ipynb: smoke-test notebook for the broad mix
• —obsidian_critic_qwen35_t4x2_unsloth_kaggle_full.ipynb: full fine-tune lineage used to produce the public LoRA run

Dataset Provenance

The training data for this first public stream comes from the mixed dataset:

• —dataset name: obsidian-critic-broad-mix-20260321
• —examples in mixed dataset: 37,008
• —approximate token volume: 6,885,960
• —exact duplicate (user, assistant) pairs removed during mix build: 3,469
• —normalized near-duplicates removed from wave backfill rows: 201
• —dataset SHA-256: 5ba1924b46d08a8ab8ad7ed5e1f74b13cc3e847b3a04b714934953975fd9300a

The public training run then created a deterministic train / validation split and applied sequence-length filtering:

• —train rows before filter: 36,638
• —validation rows before filter: 370
• —train rows after filter: 36,081
• —validation rows after filter: 363
• —removed for length filtering: 564
• —minimum kept sequence length: 48
• —maximum kept sequence length: 2048

Mix Roles

Source Dataset Table

Representative Training Examples

These are short excerpts from the actual broad-mix and component datasets used in the first public stream.

Broad mix: grounded repo repair

System

Ground the answer in one active boundary, not in general advice.

User

First fix failed. Repo: packages/notes-index
Patch 1 removed one fallback in `app/config.py`.
New evidence: `docker compose: service worker exited with code 1`.
Need the second minimal fix.

Assistant

The first patch hit the wrong seam. The new signal points back to `app/config.py`.

Broad mix: Russian ambiguity tightening

System

Пиши как строгий русскоязычный логический редактор: коротко, ясно, без филологии ради филологии.

User

Оцени формулировку из инструкции:
`Разрешён доступ аналитики и владельцы без резервной копией.`

Assistant

Да. Неясно, относится `без резервной копией` к обеим группам или только ко второй.
Строже: `И аналитики, и владельцы получают доступ без резервной копией.`

Tool-use source: strict JSON action selection

System

Ты code-workspace агент. Возвращай только валидный JSON.

User

Исправь этот сломанный tool call и верни только валидный JSON.
{"tool":"search_codebase","args":{"query":"normalize_note_path","path":"src","limit":"4"}}

Assistant

{"tool":"search_codebase","arguments":{"query":"normalize_note_path","path":"src"}}

Docs source: structured README generation

System

Ты technical writer для engineering workflows. Только полезная структура.

User

Напиши полноценный README.md для `md-audit`.
Контекст: scan Markdown repositories for broken links, empty headings, duplicate slugs, and metadata regressions.

Logic source: contradiction checking

System

Будь аккуратен с условиями и кванторами. Предпочитай точность эффектности.

User

Проверь набор на непротиворечивость.
1. `Борис` гость -> не может публиковать.
2. `Борис` гость.
3. `Борис` может публиковать.

Assistant

Нет. Противоречие между выводом из 1+2 и пунктом 3.

Training Recipe

The public run used:

• —distributed setup: `torchrun` DDP
• —training framework: Unsloth + TRL
• —base model loading: 4-bit
• —LoRA rank: 16
• —LoRA alpha: 16
• —LoRA dropout: 0.0
• —max sequence length: 2048
• —per-device train batch size: 1
• —gradient accumulation steps: 8
• —effective global batch size: 16 examples / optimization step
• —epochs: 1
• —optimizer: `adamw_8bit`
• —scheduler: cosine
• —learning rate: 1e-4
• —warmup steps: 5
• —gradient checkpointing: enabled
• —FP16: forced
• —packing: disabled
• —completion-only loss: disabled
• —public run total steps: 2256
• —logging / eval / save cadence: 50 / 125 / 250

Prompt Style

This adapter was trained on a simple, explicit prompt layout:

System:
<system prompt>

User:
<user prompt>

Assistant:

For best results, keep prompts concise, grounded, and task-shaped. The adapter responds best to:

• —repo repair tasks with concrete evidence
• —exact wording / logic cleanup tasks
• —tool-call selection with explicit schemas
• —technical writing with clear requested sections
• —review / integration prompts that specify files, symptoms, and expected outcomes

Quick Start

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

base_id = "Qwen/Qwen3.5-4B-Base"
adapter_id = "Starred09/obsidian-critic-qwen35-4b-base-lora"

tokenizer = AutoTokenizer.from_pretrained(adapter_id)
base_model = AutoModelForCausalLM.from_pretrained(
    base_id,
    torch_dtype="auto",
    device_map="auto",
)
model = PeftModel.from_pretrained(base_model, adapter_id)

system = "Return the smallest useful answer. Do not invent missing evidence."
user = "Repo: apps/desktop-shell. Build fails with ENOENT on dist/server.js. Point to the first file to inspect."
prompt = f"System:\\n{system}\\n\\nUser:\\n{user}\\n\\nAssistant:\\n"

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
with torch.no_grad():
    out = model.generate(**inputs, max_new_tokens=160)

print(tokenizer.decode(out[0], skip_special_tokens=True))

Intended Use

Graphite 1.0 4B is intended for:

• —coding assistants
• —repo triage and patch-planning copilots
• —Markdown / docs tooling assistants
• —logic and wording critique
• —bilingual technical task routing

It is especially useful when you want short, grounded, non-theatrical outputs instead of generic assistant prose.

Limitations

• —This is an adapter, not a standalone merged model.
• —It is tuned for structured technical work, not general consumer chat.
• —It inherits both strengths and weaknesses from Qwen/Qwen3.5-4B-Base.
• —The broad mix is intentionally heavy on repair, tool-use, and reasoning, so purely creative behavior is not a target.

License

This repository is released under Apache License 2.0. See `LICENSE`.

Please also review the license and usage terms of the base model:

• —`Qwen/Qwen3.5-4B-Base`

Acknowledgements

• —Alibaba Qwen team for the base model
• —Unsloth for the efficient LoRA training stack
• —TRL / Transformers / PEFT / PyTorch maintainers