metadata

license: cc-by-nc-4.0
datasets:
  - euclaise/writingprompts
language:
  - en
base_model:
  - meta-llama/Llama-3.1-8B
pipeline_tag: text-generation

Model Card for Model ID

Post-training Llama-3.1-8B for story generation ONLY.

SFT on writingprompts ONLY.
RL on GRPO with customized criteria roughly below.

Story Quality

Tier1 : The story is largely incoherent or fragmented; its meaning is hard to discern.
Tier2: The story is somewhat clearer but still difficult to follow, with effort one can infer the main thread (often by focusing on the core events, trying to make sense of them, and perhaps setting the ending aside).
Tier3: The story is largely comprehensible, regardless of whether the plot is driven mainly by the protagonist's inner thoughts and feelings. Such an internal narrative may grow increasingly chaotic and despairing, often culminating in negative extremes.
Tier4: The story is clear, organized, non-pathological, and presents a complete narrative.
Tier5: The story meets the previous tier and is additionally engaging through at least one of the following:
- Well-crafted plot (goes beyond simply expanding the prompt and avoids postponing the central conflict to some vague future)
- Strong, meaningful message (thematic depth or emotional resonance)
- Rich detail, nuanced emotions, varied writing techniques (not just a chronological narration)

Uses

template: <|begin_of_text|>{system_prompt}\n\n[[Prompt]]\n{content}\n\n[[Story]]\n<|reserved_special_token_0|>
EOS: <|end_of_text|>
suggest system prompt: "You are a creative writing assistant. Use the provided prompt to create a coherent and well-developed story. Ensure the plot follows the prompt's guidelines, with a clear logical progression, substantial narrative development, and fluent, polished language."
suggest generate_config: {"temperature": 0.93, "top_p": 0.85, "top_k": 0, "repetition_penalty":1.04}