Hostile Neural Networks guide for FTB Evolution 2 1.0.2 to 1.3.1

Hostile Neural Networks lets you simulate mobs to get their drops without hunting them. It is based on Deep Mob Learning. You train data models, run simulations with them, and turn the resulting predictions into mob loot. This pack ships version 7.0.0 of the mod.^[1][2]

How the loop works

The core idea is the Data Model, an item that holds information about one type of entity. To get one, you craft a Data Model Framework and use it on the target entity to build the model. After that, you have to train it.^[1]

1. Build a Data Model from a Data Model Framework by using the framework on the target entity.^[1]
2. Train the model. You can do this with a Deep Learner or with a Simulation Chamber. The Deep Learner holds a few models, and the kills you make begin to train them. Every model starts out Faulty and has to be trained up to Basic in the Deep Learner.^[1]
3. Once the model is Basic or higher, run simulations in the Simulation Chamber. This needs energy (based on the model type) and a catalyst item, which is usually a Prediction Matrix.^[1]
4. Feed the Mob Predictions into a Loot Fabricator. It uses power to turn predictions into mob drops.^[1]

Simulations always give a generalized prediction, and they may also give a Mob Prediction depending on how accurate the model is. Higher model tiers are more accurate, up to 99.5% at Self Aware.^[1]

Custom models

Data Models can also be added with a datapack. Each model is a single file in the data_models subfolder.^[1]

Hostile Neural Networks lets you simulate mobs and turn the results into real drops. The loop is simple: build a model of a mob, train it, run simulations with it, then turn the predictions into loot. Here is the order to make things in.^[1]

Step one: make a Model Framework

Everything starts with the Model Framework. You use it on the mob you want to copy, and it becomes a Data Model for that mob. [[config:config/hostilenetworks.cfg#models.Right Click To Attune]] is on in this pack, so right clicking a blank model on a mob or block attunes it. The pack changed this recipe.^[1][3][4]

Model FrameworkChanged by the pack

Crafting Shaped

Crafting Shaped

any silicon	any basic circuit	any silicon
any redstone dust	Smooth Stone	any redstone dust
any silicon	any gold ingot	any silicon

Recipe id: hostilenetworks:framework

The pack's own Model Framework recipe.

Step two: make a Deep Learner and train the model

Next, craft the Deep Learner. It holds a few data models, and kills you make begin to train them. Every model starts out Faulty and must be trained up to Basic here before it is useful for simulations. With [[config:config/hostilenetworks.cfg#models.Action Upgrades Model]] on, killing mobs and breaking blocks upgrade the matching model. The pack changed this recipe too.^[1][3][4]

Deep LearnerChanged by the pack

Crafting Shaped

Crafting Shaped

any obsidian	any basic circuit	any obsidian
any basic circuit	any glass panes	any basic circuit
any obsidian	any redstone dust	any obsidian

Recipe id: hostilenetworks:deep_learner

The pack's own Deep Learner recipe.

Step three: build a Simulation Chamber

Once a model is Basic or higher, the Simulation Chamber can run simulations with it. Every simulation produces a generalized prediction, and it may also produce a Mob Prediction. The higher the model's tier, the better the accuracy, up to 99.5% at Self Aware. The chamber stores up to 2,000,000 FE, set by [[config:config/hostilenetworks.cfg#power.Sim Chamber Power Cap]].^[1][3]

Simulation ChamberChanged by the pack

Crafting Shaped

Crafting Shaped

Empty slot	any glass panes	Empty slot
Ender Pearl	Machine Block	Ender Pearl
any steel plate	any basic circuit	any steel plate

Recipe id: hostilenetworks:sim_chamber

The pack changed this recipe.

Simulations also need energy and a catalyst item, which is usually the Prediction Matrix. The pack's recipe makes 16 at a time.^[1][4]

16× Prediction MatrixChanged by the pack

Crafting Shaped

Crafting Shaped

any iron ingot	any glass panes	Empty slot
any glass panes	any silicon	any glass panes
Empty slot	any glass panes	any gold ingot

Recipe id: hostilenetworks:prediction_matrix

The chamber also has a Training mode, which raises a model's data through simulation. [[config:config/hostilenetworks.cfg#models.Sim Chamber Upgrades Model]] is set to 1 (yes) in this pack. The screen text says the resources for training come from a Data Center.^[3][5]

Step four: turn predictions into drops

Mob Predictions are processed in a Loot Fabricator, which uses power to turn them into mob drops. You pick a target for each type of model, and it remembers your choices. In this pack it costs 256 FE/t to run, the same as the mod default.^[1][3][5]

Step five: upgrade to the Data Center

When you want to scale up, the Data Center runs up to 25 Data Models at once. The pack changed its recipe, and also the recipe for the Data Center IO Port.^[4][5]

Data CenterChanged by the pack

Crafting Shaped

Crafting Shaped

Echo Shard	Nether Star	Echo Shard
Simulation Chamber	any advanced circuit	Simulation Chamber
any iridium plate	Advanced Machine Block	any iridium plate

Recipe id: hostilenetworks:data_center

Data Center IO PortChanged by the pack

Crafting Shaped

Crafting Shaped

any steel plate	Black Stained Glass	any steel plate
Black Stained Glass	any basic circuit	Black Stained Glass
any steel plate	Black Stained Glass	any steel plate

Recipe id: hostilenetworks:data_center_io_port

IO Ports can replace any wall or ceiling block of the shell.

The Data Center needs a hollow 7×7×7 shell built around it:^[5]

• The floor must be obsidian.^[5]
• The walls and ceiling must be black stained glass.^[5]
• The controller replaces one bottom-wall block and must face outwards.^[5]
• Right click an IO Port to cycle its mode between Energy, Models, Inputs and Outputs.^[5]

The pack tuned the Data Center's power. Its storage, [[config:config/hostilenetworks.cfg#power.Data Center Power Cap]], is 20,000,000 FE here (the mod default is 8,000,000). The per-model cost multiplier, [[config:config/hostilenetworks.cfg#models.Data Center Sim Cost Multiplier]], is 1.25 instead of the mod default of 1.5.^[3]

Hostile Neural Networks revolves around the Data Model, an item that holds information about one entity so you can turn it into that entity's drops. The whole chain runs from a Model Framework, through training and simulation, to a Loot Fabricator. Here is how each step works in this pack.^[1]

From framework to drops

1. Craft a Model Framework and use it on the target entity to build the model. In this pack right-click attuning is switched on, so using a blank model on a mob or block attunes it to that target.^[1][3]
2. Train the model. A Deep Learner holds a few models, and the kills you make train them. Every model starts Faulty and has to be trained up to Basic in the Deep Learner.^[1]
3. Once a model is Basic or higher, run it in the Simulation Chamber. A simulation always produces a generalized prediction and may also produce a Mob Prediction, depending on the model's accuracy. It needs energy and a catalyst item, which is usually a Prediction Matrix.^[1]
4. Feed the Mob Predictions into a Loot Fabricator. It spends power to turn them into drops from a predefined list. You pick a target for each model type, and the fabricator remembers your choice for each type, so one fabricator can take several kinds of prediction.^[1][5]

Accuracy rises with the model's tier, up to 99.5% at Self Aware.^[1]

Training and accuracy

Killing mobs and breaking blocks upgrades the data on the matching model, and [[config:config/hostilenetworks.cfg#models.Action Upgrades Model]] is on. [[config:config/hostilenetworks.cfg#models.Continuous Accuracy]] is also on, so a model's accuracy keeps climbing as it progresses toward the next tier, instead of staying at the base accuracy of its current tier.^[3]

The Simulation Chamber can run in inference mode or training mode. In inference mode, simulations produce output but the model gains no data. In training mode, simulations increase the model's data, and in this pack the resources for that come from a Data Center. [[config:config/hostilenetworks.cfg#models.Sim Chamber Upgrades Model]] is set to 1, meaning training mode upgrades models normally. At 0 it would be disabled and take training mode with it, and at 2 it would only upgrade up to tier boundaries.^[3][5]

The pack also turns on [[config:config/hostilenetworks.cfg#models.Enable Block Data Models]], which the mod leaves off by default. It loads block data models such as the built-in ore models. The config describes it as an experimental feature.^[3]

Power

These are the FE buffers and costs this pack uses:^[3]

• Simulation Chamber buffer: 2,000,000 FE ([[config:config/hostilenetworks.cfg#power.Sim Chamber Power Cap]]), the mod default.^[3]
• Loot Fabricator: a 1,000,000 FE buffer ([[config:config/hostilenetworks.cfg#power.Loot Fab Power Cap]]) and a cost of 256 FE/t ([[config:config/hostilenetworks.cfg#power.Loot Fab Power Cost]]), both mod defaults.^[3]
• Data Center buffer: 20,000,000 FE ([[config:config/hostilenetworks.cfg#power.Data Center Power Cap]]). The pack raised this from the mod default of 8,000,000, and the config notes it is sized for 25 concurrent self-aware simulations.^[3]
• Data Center running cost: [[config:config/hostilenetworks.cfg#models.Data Center Sim Cost Multiplier]] is 1.25 in this pack, down from the mod default of 1.5. It multiplies each active model's base simulation cost, paid per tick for every active model.^[3]

The Data Center multiblock

The Data Center runs up to 25 Data Models at the same time. Each one produces one drop per cycle into a shared output buffer. In this pack it is also what supplies the resources for Simulation Chamber training.^[5]

It needs a hollow 7×7×7 shell. The floor must be obsidian, and the walls and ceiling must be black stained glass. The controller replaces one bottom-wall block and must face outwards.^[5][6]

Besides vanilla black stained glass, the wall tag also accepts some black glass variants from other mods, but only when those mods are installed. These include the black glass types from Connected Glass and the reinforced black stained glass from SecurityCraft.^[7]

Data CenterChanged by the pack

Crafting Shaped

Crafting Shaped

Echo Shard	Nether Star	Echo Shard
Simulation Chamber	any advanced circuit	Simulation Chamber
any iridium plate	Advanced Machine Block	any iridium plate

Recipe id: hostilenetworks:data_center

The pack changed this recipe from the mod's own.

To operate it, put Data Models into the 5×5 grid and matching items into the four input slots. Each active slot uses up one input per cycle and produces the drop chosen by its Fabrication Directive.^[5]

IO Ports and automation

The Data Center IO Port can replace any wall or ceiling block of the shell, and the controller detects it automatically. Right-click it to cycle through four modes: Energy, Models, Inputs and Outputs. Energy accepts power, while Models, Inputs and Outputs expose the relevant item slots, so you can automate those through the shell.^[5][7]

Data Center IO PortChanged by the pack

Crafting Shaped

Crafting Shaped

any steel plate	Black Stained Glass	any steel plate
Black Stained Glass	any basic circuit	Black Stained Glass
any steel plate	Black Stained Glass	any steel plate

Recipe id: hostilenetworks:data_center_io_port

The pack changed this recipe too.

To set up another fabricator the same way, use a Fabrication Directive. It stores the selections from a Loot Fabricator. Right-click to copy a layout, and Shift-Right-click to apply it.^[5]

Predictions

Generalized predictions come in Overworld, Nether, End and Twilight versions. The pack has crafting recipes for the Generalized Nether Prediction and the Generalized Ender Prediction.^[4][8]

Generalized Nether PredictionMod recipe

Crafting Shaped

Crafting Shaped

Empty slot	Overworld Prediction	Empty slot
Overworld Prediction	Netherrack	Overworld Prediction
Empty slot	Overworld Prediction	Empty slot

Recipe id: hostilenetworks:living_matter/overworld_to_nether

Generalized Ender PredictionMod recipe

Crafting Shaped

Crafting Shaped

Empty slot	Generalized Nether Prediction	Empty slot
Generalized Nether Prediction	End Stone	Generalized Nether Prediction
Empty slot	Generalized Nether Prediction	Empty slot

Recipe id: hostilenetworks:living_matter/nether_to_ender

The pack's recipe for the Prediction Matrix makes 16 per craft.^[4]

16× Prediction MatrixChanged by the pack

Crafting Shaped

Crafting Shaped

any iron ingot	any glass panes	Empty slot
any glass panes	any silicon	any glass panes
Empty slot	any glass panes	any gold ingot

Recipe id: hostilenetworks:prediction_matrix

Deep Learner HUD

The Deep Learner has a HUD element whose position you can move. It is anchored at the top left with X and Y offsets of 0, all the mod defaults. The options are [[config:config/hostilenetworks.cfg#client.Deep Learner HUD Anchor Point]], [[config:config/hostilenetworks.cfg#client.Deep Learner HUD X Offset]] and [[config:config/hostilenetworks.cfg#client.Deep Learner HUD Y Offset]]. There is also a keybind called Open Deep Learner Curio.^[3][5]

Hostile Neural Networks is a crafting chain more than a tier ladder in this pack. You build a Deep Learner, a Simulation Chamber and a Data Center, then move on to the higher-dimension predictions and the extra data models from FTB Armory. Here is the order I'd follow, with the exact in-pack recipes.^[4]

Unverified The sources here show no quest chapter for this mod, so the pack's quest book gives no set order for it. The order below follows the crafting recipes instead.

First: the Model Framework and Deep Learner

Everything starts with the Model Framework and the Deep Learner. The pack changed both recipes, so don't rely on what you remember from the mod's defaults.^[4]

Model FrameworkChanged by the pack

Crafting Shaped

Crafting Shaped

any silicon	any basic circuit	any silicon
any redstone dust	Smooth Stone	any redstone dust
any silicon	any gold ingot	any silicon

Recipe id: hostilenetworks:framework

The pack changed this recipe.

Deep LearnerChanged by the pack

Crafting Shaped

Crafting Shaped

any obsidian	any basic circuit	any obsidian
any basic circuit	any glass panes	any basic circuit
any obsidian	any redstone dust	any obsidian

Recipe id: hostilenetworks:deep_learner

The pack changed this one too.

Next: the machines

Next come the Simulation Chamber and the Data Center. The pack changed both recipes, and the Data Center IO Port and Prediction Matrix as well.^[4]

Simulation ChamberChanged by the pack

Crafting Shaped

Crafting Shaped

Empty slot	any glass panes	Empty slot
Ender Pearl	Machine Block	Ender Pearl
any steel plate	any basic circuit	any steel plate

Recipe id: hostilenetworks:sim_chamber

Changed by the pack.

Data CenterChanged by the pack

Crafting Shaped

Crafting Shaped

Echo Shard	Nether Star	Echo Shard
Simulation Chamber	any advanced circuit	Simulation Chamber
any iridium plate	Advanced Machine Block	any iridium plate

Recipe id: hostilenetworks:data_center

Changed by the pack.

Data Center IO PortChanged by the pack

Crafting Shaped

Crafting Shaped

any steel plate	Black Stained Glass	any steel plate
Black Stained Glass	any basic circuit	Black Stained Glass
any steel plate	Black Stained Glass	any steel plate

Recipe id: hostilenetworks:data_center_io_port

Changed by the pack.

16× Prediction MatrixChanged by the pack

Crafting Shaped

Crafting Shaped

any iron ingot	any glass panes	Empty slot
any glass panes	any silicon	any glass panes
Empty slot	any glass panes	any gold ingot

Recipe id: hostilenetworks:prediction_matrix

The pack's recipe makes 16 Prediction Matrix per craft.

The Data Center uses block tags for its floor and walls. The floor tag takes obsidian-type blocks. The wall tag takes black stained glass or the Data Center IO Port. If Connected Glass is installed, several of its black glass variants also count as walls.^[6][7]

Then: higher-dimension predictions

The mod groups its generalized predictions in one item tag: Overworld, Nether, End and Twilight. In this pack you can craft the Nether and End ones with the mod's own recipes, which the pack did not change.^[4][8]

Generalized Nether PredictionMod recipe

Crafting Shaped

Crafting Shaped

Empty slot	Overworld Prediction	Empty slot
Overworld Prediction	Netherrack	Overworld Prediction
Empty slot	Overworld Prediction	Empty slot

Recipe id: hostilenetworks:living_matter/overworld_to_nether

Makes the Generalized Nether Prediction. This is the mod's own recipe.

Generalized Ender PredictionMod recipe

Crafting Shaped

Crafting Shaped

Empty slot	Generalized Nether Prediction	Empty slot
Generalized Nether Prediction	End Stone	Generalized Nether Prediction
Empty slot	Generalized Nether Prediction	Empty slot

Recipe id: hostilenetworks:living_matter/nether_to_ender

Makes the Generalized Ender Prediction. This is the mod's own recipe.

Finally: FTB Armory data models

FTB Armory adds recipes of its own type, data_model_tier, that produce a Data Model: adamantite, aeternium and aurichalcum. They are mod recipes, not changes to a Hostile Neural Networks recipe.^[4]

Data ModelMod recipe

Data Model Tier (ftbarmory)

Recipe id: ftbarmory:hostilenetworks/adamantite_data_model

Data ModelMod recipe

Data Model Tier (ftbarmory)

Recipe id: ftbarmory:hostilenetworks/aeternium_data_model

Data ModelMod recipe

Data Model Tier (ftbarmory)

Recipe id: ftbarmory:hostilenetworks/aurichalcum_data_model

Same as in the FTB Skies 2: Aero 1.10.0 guide

Everything here lives in config/hostilenetworks.cfg. This pack changes only a few values, and all of them are about the Data Center and the models it can load. The rest sits at the mod defaults, so you don't need to go digging unless you want to tweak things.^[9]

What the pack changed

• [[config:config/hostilenetworks.cfg#power.Data Center Power Cap]] is 20,000,000 in this pack, up from the mod default of 8,000,000. It is the maximum FE the Data Center can store, and the config note says it is sized for 25 concurrent self-aware simulations.^[9]
• [[config:config/hostilenetworks.cfg#models.Data Center Sim Cost Multiplier]] is 1.25 in this pack, down from the mod default of 1.5. It multiplies the base simCost of each active model to set the per-tick power the Data Center pays, so running models costs a little less here than in the default setup. The allowed range is 0.01 to 100.0.^[9]
• [[config:config/hostilenetworks.cfg#models.Enable Block Data Models]] is true in this pack, while the mod default is false. This loads block data models, such as the built-in ore models. The config calls it an experimental feature, and the value is not synced because it governs which models load during datapack reading.^[9]

Power settings left at default

• [[config:config/hostilenetworks.cfg#power.Sim Chamber Power Cap]] is 2,000,000 FE, the maximum stored in the Simulation Chamber.^[9]
• [[config:config/hostilenetworks.cfg#power.Loot Fab Power Cap]] is 1,000,000 FE, the maximum stored in the Loot Fabricator.^[9]
• [[config:config/hostilenetworks.cfg#power.Loot Fab Power Cost]] is 256 FE/t, the running cost of the Loot Fabricator.^[9]

Model training settings left at default

• [[config:config/hostilenetworks.cfg#models.Action Upgrades Model]] is true, so killing mobs and breaking blocks upgrades the data on the matching model.^[9]
• [[config:config/hostilenetworks.cfg#models.Right Click To Attune]] is true, so right clicking a blank data model on a mob or block attunes it to that target.^[9]
• [[config:config/hostilenetworks.cfg#models.Continuous Accuracy]] is true, so a model's accuracy rises as it gains progress toward the next tier instead of staying at the base accuracy of its current tier.^[9]
• [[config:config/hostilenetworks.cfg#models.Sim Chamber Upgrades Model]] is 1, meaning the Simulation Chamber's Training Mode upgrades model data. The options are 0 (disabled, which also turns off Training Mode entirely), 1 (yes) and 2 (only up to tier boundaries).^[9]

Deep Learner HUD

The Deep Learner HUD is a client-side setting. It uses [[config:config/hostilenetworks.cfg#client.Deep Learner HUD Anchor Point]] set to "top_left", with [[config:config/hostilenetworks.cfg#client.Deep Learner HUD X Offset]] and [[config:config/hostilenetworks.cfg#client.Deep Learner HUD Y Offset]] both at 0, which are the mod defaults. Both offsets accept values from -1000 to 1000, so you can nudge the HUD wherever you like.^[9]

FTB Evolution 2 1.1.0 ships Hostile Neural Networks version 7.0.0, using the file HostileNeuralNetworks-26.1.2-7.0.0.jar. The mod was already in the pack as of 1.0.2 or earlier, so it is not a recent addition.^[2]

Unverified The sources here show no pack changelog entries for Hostile Neural Networks, and they do not include the text of the mod's own changelog. What changed in the mod between pack releases is therefore not confirmed in this guide. For the details, check the mod's own changelog.