TinnyLlamaGambling

A compact instruction-following model focused on responsible, factual gambling topics: RTP, house edge, expected value (EV), bankroll management, dice/coin probabilities, and Responsible Gambling guidance. Data is fully synthetic, generated with local LLMs via the provided scripts, cleaned and shuffled, then fine-tuned with LoRA and merged into a standalone model.

Model

  • Architecture: TinyLlama 1.1B and/or Mistral 7B with LoRA
  • Language: English
  • Task: text generation / instruction tuning (short, structured answers)

Data

  • Source: synthetic JSONL with fields instruction, input, output
  • Generation: sampling-controlled
  • Cleaning: trim, drop empty/too-long, shuffle

Training

  • Method: TRL SFT + PEFT LoRA; cosine LR, short epochs, seq len β‰ˆ 1k
  • LoRA targets: q_proj,k_proj,v_proj,o_proj,gate_proj,up_proj,down_proj
  • Merge: merge_lora to produce a standard Transformers checkpoint

Intended Use

  • Educational explanations of gambling math and concepts (EV, RTP, house edge)
  • High-level bankroll/risk topics and Responsible Gambling reminders
  • Not financial/betting advice

Limitations

  • Synthetic data may contain errors or simplifications
  • Do not use outputs to make wagering decisions
  • Avoid encouraging gambling or irresponsible behavior

Inference (Transformers)

from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "galanviorel/TinnyLlamaGambling"
tok = AutoTokenizer.from_pretrained(model_id, use_fast=True)
mdl = AutoModelForCausalLM.from_pretrained(model_id)

prompt = "### Instruction:\nExplain house edge for a game with RTP=96%.\n\n### Response:\n"
x = tok(prompt, return_tensors="pt")
y = mdl.generate(**x, max_new_tokens=128, temperature=0.7, top_p=0.9)
print(tok.decode(y[0], skip_special_tokens=True))

Recommended Generation Params

  • temperature: 0.6–0.9
  • top_p: 0.9–0.95
  • repetition_penalty: 1.05–1.15
  • max_new_tokens: 128–256

License

  • Model card/code: Apache-2.0
  • Base models: per their original licenses
  • Data: synthetic; follow local laws and Responsible Gambling guidelines
Downloads last month
32
Safetensors
Model size
1B params
Tensor type
BF16
Β·
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for galanviorel/TinnyLlamaGambling

Adapter
(1271)
this model