---
library_name: transformers
license: apache-2.0
base_model: Qwen/Qwen3.6-27B
tags:
  - graphforge
  - supervised-fine-tuning
  - tool-use
  - working-agents
---

# GraphForge-Qwen3.6-27B-SFT

This is a final GraphForge SFT checkpoint for file-based working agents,
fine-tuned from Qwen/Qwen3.6-27B. This is **not** an RFT model.

## Training overview

| Item | Setting |
|---|---|
| Training corpus | GraphForge-SFT-2169 |
| Training examples | 2,169 trajectory sequences |
| Epochs | 3 |
| Data admission | Teacher agent-judge score strictly above 0.90 |
| Reasoning history | Preserved in the training trajectories |
| Sequence-length policy | Discard complete samples above 262,144 tokens rather than truncate |

The GraphForge 27B and 35B-A3B SFT releases use the same main training corpus.
The data pipeline constructs tasks and evidence-anchored rubrics from real files,
uses an initial execution and one revision stage, and exports admitted teacher
trajectories for supervised learning. The teacher model is GLM-5.2.

## Contents and inference

The repository contains the final safetensors weights, the weight index,
model configuration, tokenizer, chat template, and available processor assets.
It excludes optimizer state, training logs, and intermediate checkpoints.
Use the included chat template with a compatible Transformers or serving runtime.
The exported architecture identifier may use Qwen3.5-compatible implementation
names; do not rename architecture fields merely to match the release name.

Keep reasoning history when running multi-turn tool interactions. Provide the
tools and workspace required by the task; weights alone do not supply an agent
runtime, web credentials, OCR service, or document-processing environment.

## Evaluation and limitations

Benchmark results depend on the scaffold, available tools, task coverage, and
judge protocol. Internal GDPVal Elo values are not interchangeable with official
leaderboard scores. This card does not substitute a single score for those settings.
The model can produce incorrect facts, unsafe tool actions, or invalid files.
Review outputs and use appropriate sandboxing before executing generated commands.

## Attribution and license

This model is derived from Qwen/Qwen3.6-27B. The upstream Apache 2.0 license
is retained in LICENSE. The weights were fine-tuned, this model card was replaced,
and local provenance fields were removed from release configurations where present.
This repository is the GraphForge derivative, not the unmodified upstream model.