Escarda-86M-Identity — Anti-DEG
A tuned build of Escarda-86M-Identity (86M parameters, SpikeWhaleLM) that
follows a change of subject in conversation while keeping the Escarda identity.
Ships as a fully merged standalone model — no adapters or base model needed.
Why this build exists
The Escarda-86M-Base-JL-Anti-DEG build scores higher on factual questions but has
no persona — asked "who are you" it answers "I am a young woman." This build
applies the same treatment to the identity-tuned checkpoint instead, so the model
still knows what it is:
"Who are you?" → "I'm Escarda, a helpful AI assistant created by Quazim0t0 (Dean Byrne)."
Scores
Loops measured on 6 prompt stems; topic adherence on 6 held-out multi-turn topic switches; factual accuracy on 73 unambiguous questions. All at effort n=2 with a 24-token minimum length.
Escarda-86M-Identity | this model | |
|---|---|---|
| topic adherence (1.0 = fully follows a subject change) | 0.75 | 1.00 |
| topic switches answered on the old subject | 0 / 6 | 0 / 6 |
| repetition loops | 0 / 6 | 0 / 6 |
| answers as "Escarda" | 2 / 2 | 2 / 2 |
| factual accuracy (73 items) | 0.110 ± 0.037 | 0.068 ± 0.030 |
Read the factual row honestly: the difference is −0.042 with a 0.048 standard
error on the difference — inside one standard error, i.e. not a measurable change.
This build buys topic adherence at no detectable factual cost; it does not improve
factual accuracy. For that, Escarda-86M-Base-JL-Anti-DEG scores 0.260 ± 0.051
(2.4σ better) but loses the persona entirely.
At 86M parameters this model is not a source of reliable facts either way.
Decoding matters
The scores above are measured with effort decoding (effort_decode.py,
included) plus a 24-token minimum reply length. Greedy decoding does not reach
them. Effort levels trade compute for quality: low draws a single sample, high
draws six and returns the best.
python effort_decode.py --model Quazim0t0/Escarda-86M-Identity-Anti-DEG \
--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
Usage
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
tok = AutoTokenizer.from_pretrained(
"Quazim0t0/Escarda-86M-Identity-Anti-DEG", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
"Quazim0t0/Escarda-86M-Identity-Anti-DEG",
trust_remote_code=True, dtype=torch.float32).eval()
Prompt format is ChatML:
<|im_start|>user
{question}<|im_end|>
<|im_start|>assistant
Limitations
- 86M parameters, and not instruction-tuned for factual accuracy. Output is often confidently wrong; do not rely on it for facts.
- Native 4096-token context (no Jet-Long on this checkpoint).
- Custom architecture and tokenizer —
trust_remote_code=Trueis required.