hanifdev's picture
Model card
122375b verified
|
Raw History Blame Contribute Delete
2.35 kB
---
license: apache-2.0
base_model:
- Qwen/Qwen3-4B-Instruct-2507
- Qwen/Qwen3-8B
tags:
- gguf
- llama.cpp
- lora
- game-development
---
# Taygeta Dream event models
Fine-tunes that write event data for the game **Taygeta Dream** from a plain-language request. They are the
local backends of the `taygeta_events` generator, which supplies the prompt (the game's exported schema and a
catalog of its triggers, actions, and requirements), validates every answer, and asks the model to repair it.
On their own, without that prompt, they aren't useful as general assistants.
| File | Tier | What it is |
|---|---|---|
| `qwen3-4b-taygeta-ft4-Q4_K_M.gguf` (2.5 GB) | 4 GB GPU | Qwen3-4B-Instruct-2507 + the ft4 QLoRA, merged and quantized |
| `qwen3-8b-taygeta-ft4-Q4_K_M.gguf` (5.0 GB) | 8 GB GPU | Qwen3-8B + the ft4 QLoRA, merged and quantized. Serve with thinking off |
| `qwen3-4b-taygeta-ft4-lora-f16.gguf` (66 MB) | 4 GB GPU | The 4B ft4 adapter alone, for llama.cpp `--lora` on the untrained base [`unsloth/Qwen3-4B-Instruct-2507-GGUF`](https://huggingface.co/unsloth/Qwen3-4B-Instruct-2507-GGUF) `Qwen3-4B-Instruct-2507-Q4_K_M.gguf`. With the adapter at scale 0 the same server plans long requests as the base model; at scale 1 it writes like the merged file |
## Serving
llama.cpp `llama-server`, as the game starts it:
```
llama-server -m qwen3-4b-taygeta-ft4-Q4_K_M.gguf -c 20480 -np 1 -fa on -ctk q4_0 -ctv q4_0 --dry-multiplier 0.8
llama-server -m qwen3-8b-taygeta-ft4-Q4_K_M.gguf -c 20480 -np 1 -fa on -ctk q8_0 -ctv q8_0 --dry-multiplier 0.8 \
--chat-template-kwargs '{"enable_thinking":false}'
```
The DRY sampler matters: without it the fine-tunes can repeat a dialog line until the token limit.
## Results
Share of evaluation requests answered with valid events that do what was asked, with the generator's grammar,
validator and repair loop (55 general requests / 36 requests from a second author):
| Model | General | Second author |
|---|---|---|
| 4B ft4 | 82.7% | 77.8% |
| 8B ft4 | 90.0% | 88.9% |
## Training
QLoRA on ~1,800 request → events examples: back-translated requests for the game's own event chains,
schema-sampled events, refusal and contrast examples for off-task requests, and targeted patterns
(thresholds, gifts, purchases, choices). Requests used for evaluation are held out.