Phips/qwen3.5-coding-finetune-code

Qwen3.5/3.6 Agentic Coding Finetune

Produce releasable GGUF models for Kilo Code / OpenCode that emit correct tool calls and avoid Ornith loop/broken-tool-call failures.

Goal

Train an open-source coding agent that works reliably in local IDEs (Kilo/OpenCode/Omp.sh) via XML tool calling. The 4B model trains entirely on consumer hardware (RTX 3060), then GRPO scales it to 27B.

Pipeline: SFT → GRPO → Release

Stage	Model	Hardware	Data	Purpose
V1	Phhofm/qwen3.5-4b-agentic-local	RTX 3060 12GB	8.5K canonical tool traces	Prove pipeline viability locally
V2	Phhofm/qwen3.5-4b-agentic-v2	Vast.ai A100 80GB	V1 + multi-turn GRPO	RL-hardened recovery, anti-loop
V3	Phhofm/qwen3.6-27b-agentic	Vast.ai	V2 scaled to 27B	Beat Ornith on SWE-bench

Current Status

Training in progress: V1 SFT on RTX 3060

• 8.5K canonical tool traces (all 10 tools)
• LoRA r=32, batch=2, grad_accum=4
• Checkpoint available: outputs/sft_v1/checkpoint-500/

Canonical Tools (10)

The model emits XML tool calls compatible with Kilo/OpenCode:

<read><path>file.py</path></read>
<edit><path>file.py</path><old_string>x</old_string><new_string>y</new_string></edit>
<bash><command>ls -la</command></bash>
<glob><pattern>**/*.py</pattern></glob>
<grep><pattern>TODO</pattern></grep>
<write><path>out.py</path><content>def f(): pass</content></write>
<websearch><query>python requests api</query></websearch>
<webfetch><url>https://docs.python.org/3/</url></webfetch>
<question><questions>What file should be modified?</questions></todowrite>

Omp.sh compatibility: Uses XML tool calling. Likely compatible with same format. Check their docs at omp.sh for exact param names.

Quick Start

# Setup
uv venv .venv --python 3.12
uv pip install -r requirements.txt

# V1 SFT (local)
./train/run_safe.sh sft_4b --data data/v5_training.jsonl --context 4096 --lora-r 32

# Validate (when training complete)
./train/run_safe.sh post_sft_validate --model outputs/sft_v1 --tasks benchmark/benchmark_tasks.jsonl

# Convert to GGUF
.venv/bin/python train/release.py --model outputs/sft_v1 --output models/

Data

File	Description
data/v5_tool_patterns.jsonl	Converted tool traces (4.8K rows)
data/v5_synthetic.jsonl	Recovery/grounding trajectories (25K rows)
data/v5_training.jsonl	Merged training set (8.5K unique rows)

Key Files

File	Purpose
IMPLEMENTATION.md	Full pipeline plan with milestones
train/sft_4b.py	SFT training with QLoRA
train/grpo_4b.py	GRPO training (future: ART)
train/kilo_parser.py	Canonical parser/renderer
train/post_sft_validate.py	Validation gate (§7)
train/gen_synthetic.py	Recovery trajectory generator

Inference Settings

• Temperature: 1.0 (lower causes repetition loops)
• top_p: 0.95
• KV cache: BF16