qwen3.5-9b-factorio-build-sft API & Inference Endpoint

Results

Success rate on factorio-build (8 scenes per row, 16 rows × 4 samples, temperature 0.6, thinking off):

Table with columns: Model, Validation, Holdout
Model	Validation	Holdout
Qwen3.5-9B	2.1%	1.6%
This model	18.0%	18.0%
This model + 20 GRPO steps	93.8%	79.5%
SFT, 800 programs (sft-800)	67.0%	57.4%
SFT, 800 programs + 20 GRPO steps (sft-800-grpo)	85.9%	84.0%

Sampling matters: at temperature 1.0 this model's long programs drift into syntax errors and success falls to ~0%.

Training

Data: 93 distinct programs found by LLM-guided program search (evolve in factory-sim), each scoring ≥ 0.5 on training scenes and passing the current sandbox, paired with real factorio-build training prompts. No validation or holdout scene was used for selection.
Method: LoRA (rank 32, alpha 64) on every attention, Gated DeltaNet and MLP projection of the language model, merged into the weights. Loss on the completion only; prompts rendered with the Qwen3.5 chat template and enable_thinking=False.
Schedule: 2 epochs, 24 optimizer steps (batch 8), AdamW, lr 1e-4 with warm-up and cosine decay, bf16, one A100 80GB, ~8 minutes.

Recipe, configs and scripts: recipes/grpo-factorio-build.

Use

Serve with vLLM and pass chat_template_kwargs={"enable_thinking": false}. The prompt format is the one factorio-build produces; the model answers with a short plan and one fenced Python block.

Limitations

Trained for one task (construct_smelting_line) and one API; it is not a general coding model.
Programs act only through the simulator's world API, inside a sandbox.
Numbers come from 64 samples per split; expect a few points of noise.

Results

Success rate on factorio-build (8 scenes per row, 16 rows × 4 samples, temperature 0.6, thinking off):

Table with columns: Model, Validation, Holdout
Model	Validation	Holdout
Qwen3.5-9B	2.1%	1.6%
This model	18.0%	18.0%
This model + 20 GRPO steps	93.8%	79.5%
SFT, 800 programs (sft-800)	67.0%	57.4%
SFT, 800 programs + 20 GRPO steps (sft-800-grpo)	85.9%	84.0%

Sampling matters: at temperature 1.0 this model's long programs drift into syntax errors and success falls to ~0%.

Training

Data: 93 distinct programs found by LLM-guided program search (evolve in factory-sim), each scoring ≥ 0.5 on training scenes and passing the current sandbox, paired with real factorio-build training prompts. No validation or holdout scene was used for selection.
Method: LoRA (rank 32, alpha 64) on every attention, Gated DeltaNet and MLP projection of the language model, merged into the weights. Loss on the completion only; prompts rendered with the Qwen3.5 chat template and enable_thinking=False.
Schedule: 2 epochs, 24 optimizer steps (batch 8), AdamW, lr 1e-4 with warm-up and cosine decay, bf16, one A100 80GB, ~8 minutes.

Recipe, configs and scripts: recipes/grpo-factorio-build.

Use

Serve with vLLM and pass chat_template_kwargs={"enable_thinking": false}. The prompt format is the one factorio-build produces; the model answers with a short plan and one fenced Python block.

Limitations

Trained for one task (construct_smelting_line) and one API; it is not a general coding model.
Programs act only through the simulator's world API, inside a sandbox.
Numbers come from 64 samples per split; expect a few points of noise.

qwen3.5-9b-factorio-build-sft

README

Results

Training

Use

Limitations

Explore FriendliAI today

README

Results

Training

Use

Limitations