Quazim0t0/Escarda-86M-Base-JL-Anti-DEG

Escarda-86M-Base-JL — Antidoom

A tuned build of Escarda-86M-Base-JL (86M parameters, SpikeWhaleLM) that eliminates repetition loops and follows a change of subject in conversation.

Ships as a fully merged standalone model — no adapters or base model needed.

Scores

Measured over 90 generations (30 prompt stems x 3 temperatures) for loops, and 6 held-out topic switches in multi-turn chat for topic adherence.

	base	this model
repetition loops	90% of generations	0%
topic adherence (1.0 = fully follows a subject change)	0.33	1.00
topic switches answered on the old subject	4 / 6	0 / 6
bigram diversity (distinct-2)	0.148	0.874
held-out perplexity (wikitext-2 test, 30,720 tokens)	8.22	8.72

Perplexity rises 6.0%.

Decoding matters

The loop and topic scores above are measured with effort decoding (effort_decode.py, included). Greedy decoding on this model does not reach them. Use the included script, or the same approach in your own loop.

python effort_decode.py --model Quazim0t0/Escarda-86M-Base-JL-Antidoom \
  --prompt "<|im_start|>user\nHow do I change a car tyre?<|im_end|>\n<|im_start|>assistant\n" \
  --question "How do I change a car tyre?" --level high

Effort levels trade compute for quality: low is a single sample, high draws six and returns the best. On 2 CPU threads at 120 new tokens, low is ~3s and high ~22s.

Usage

import torch
from transformers import AutoModelForCausalLM

model = AutoModelForCausalLM.from_pretrained(
    "Quazim0t0/Escarda-86M-Base-JL-Antidoom",
    trust_remote_code=True, dtype=torch.float32).eval()

Pass explicit position_ids during cached decode, as the base architecture requires. Chat format is ChatML (<|im_start|>role\n...<|im_end|>).

Limitations

• 86M parameters and not instruction-tuned for factual accuracy. Output is often wrong on facts; only repetition and topic behaviour were targeted.
• Topic adherence is measured on 6 held-out switches — enough to establish the effect, not to give a precise figure.
• The loop detector used for scoring misses slow-alternating repetition, so distinct-2 is reported alongside it.