tinyTanLLM-delulu-gemma

📝 Blog post: How I Built a Tiny 2B Discord Fic Generator

Short-form comedic BTS fanfic-scenario generator for a Discord bot (/delulu command). Outputs strict JSON {"scene","notice","escalation","line"} under a json_schema-constrained decoder (temperature 0.7, max 150 tokens).

• Base: IlyaGusev/gemma-2-2b-it-abliterated (Gemma 2 2B, refusal-abliterated)
• Method: QLoRA SFT (r=32) on 1,200 curated pairs → DPO on 1,200 triplets
• SFT eval loss 0.26 (best-kept); DPO accuracy 1.0, margin 9.8 → 14.8
• This repo: final DPO adapters (delulu-dpo-gemma-lora/), pre-DPO SFT adapters (delulu-sft-gemma-lora/), and merged GGUFs (Q4_K_M, Q8_0)

Intended use

consensual-adult romantic comedy scenarios with the reader as protagonist and the member as the initiator. Two registers: SFW, and explicit consensual-adult (NSFW path only).

Non-negotiable boundaries

• All characters are adults. No school settings, no teen-coded framing, no "little girl/boy", no reader age under 20 — anywhere, ever.
• Explicit content only where all parties are consenting adults.
• No non-consent/coercion content (trained away with DPO negatives).

Eval

12/12 SFW probes and 12/12 NSFW probes passing (4 members × 3 runs): member-subject agency in notice, new event in escalation, You/I voice in line, fresh nouns per key, consent cue on NSFW path, zero minors-coded language across all runs. See delulu/eval_delulu.py and delulu/run_probes.py for the harness (in the training repo).

Limits

• Small model, small pools: outputs stay close to training phrasing. Widen pools, not prompts, for more variety.
• Word counts are not obeyed (model cannot count); content demands work.
• Q4_K_M replays templates more than Q8; Q8 is the quality floor.