Downloads · 30 days
9
29% of all-time downloads
G4-MeroMero-v2-31B
rk35784/G4-MeroMero-v2-31B
G4-MeroMero-v2-31B is a machine learning model from rk35784. Use it for the machine learning task on the model card, and read the license before you ship it in a product. The card lists the license as apache-2.0.
<style .gs { --bg: 0f1014; --surface: 171922; --edge: 262a36; --rule: 333849; --text: b8bfcc; --dim: 7d8595; --bright: f2f4f8; --azure: ef5d76; --crimson: 5b9dff; --az-glow: rgba(91,157,255,0.08); --cr-glow: rgba(239,…
All-time downloads
31
Public
Parameters
31.3B
62.6 GB on disk
Likes
0
Public
Repo makeup
Click a slice to open those files.
.safetensors62.5 GB · 100%
At a glance
- License
- apache-2.0
- Model type
- gemma4
- Access
- Public
- Created
- Aug 17, 2026
- Updated
- Aug 17, 2026
- SHA
- 7bd01068
Base models
Model card
From the Hugging Face model README
<style>
.gs {
--bg: #0f1014;
--surface: #171922;
--edge: #262a36;
--rule: #333849;
--text: #b8bfcc;
--dim: #7d8595;
--bright: #f2f4f8;
--azure: #ef5d76;
--crimson: #5b9dff;
--az-glow: rgba(91,157,255,0.08);
--cr-glow: rgba(239,93,118,0.06);
--mono: 'JetBrains Mono', monospace;
--sans: 'Inter', sans-serif;
font-family: var(--sans);
color: var(--text);
max-width: 900px;
margin: 0 auto;
padding: 0 0 60px;
line-height: 1.7;
font-size: 1rem;
background:
radial-gradient(ellipse at 50% 0%, rgba(91,157,255,0.08) 0%, transparent 50%),
radial-gradient(ellipse at 50% 100%, rgba(239,93,118,0.05) 0%, transparent 50%),
var(--bg);
}
/* ── Profile Card ── */
.gs-profile {
border-bottom: none;
position: relative;
background: var(--surface);
margin-bottom: 0;
}
.gs-profile-art {
position: relative;
}
.gs-profile-art img {
display: block;
width: 100%;
height: 380px;
object-fit: cover;
margin-top: 0px;
}
.gs-ident {
position: absolute;
bottom: 0;
left: 0;
right: 0;
padding: 120px 44px 28px;
background: linear-gradient(
to top,
rgba(11,12,16,0.88) 0%,
rgba(11,12,16,0.5) 50%,
transparent 100%
);
}
.gs-profile-info {
padding: 20px 44px 36px;
display: flex;
flex-direction: column;
gap: 20px;
}
.gs-profile-label {
display: flex;
align-items: baseline;
gap: 10px;
font-family: var(--mono);
letter-spacing: 0.14em;
text-transform: uppercase;
}
.gs-profile-label .gs-snum {
font-size: 0.62rem;
font-weight: 700;
color: var(--crimson);
opacity: 1;
position: static;
transform: none;
}
.gs-profile-label .gs-stitle {
font-size: 0.62rem;
color: var(--dim);
font-weight: 700;
letter-spacing: 0.14em;
}
.gs-profile-label .gs-stitle::before {
content: none;
}
.gs-name {
font-family: var(--sans);
font-size: 3.2rem;
font-weight: 900;
color: #ffffff;
letter-spacing: 0.06em;
line-height: 1;
margin: 0 0 10px;
text-shadow: 0 2px 12px rgba(0,0,0,0.3);
overflow-wrap: break-word;
}
.gs-base {
font-family: var(--mono);
font-size: 0.68rem;
color: var(--crimson);
letter-spacing: 0.14em;
text-transform: uppercase;
display: block;
}
.gs-profile-bio p {
margin: 0 0 14px;
font-size: 0.95rem;
}
.gs-profile-bio p:last-child { margin-bottom: 0; }
/* ── Sections ── */
.gs-section {
padding: 0;
}
.gs-shead {
position: relative;
display: flex;
align-items: center;
gap: 14px;
padding: 16px 44px 14px;
margin-bottom: 28px;
border-top: 2px solid;
border-image: linear-gradient(90deg, var(--crimson), var(--azure)) 1;
}
.gs-snum {
font-family: var(--mono);
font-size: 2.2rem;
font-weight: 900;
color: var(--crimson);
letter-spacing: 0.06em;
opacity: 0.12;
position: absolute;
right: 44px;
top: 50%;
transform: translateY(-50%);
line-height: 1;
}
.gs-stitle {
font-size: 1.05rem;
font-weight: 700;
letter-spacing: 0.1em;
text-transform: uppercase;
color: var(--bright);
}
.gs-stitle::before {
content: '\2726';
color: var(--crimson);
font-size: 0.8em;
margin-right: 8px;
}
.gs-sbody {
padding: 0 44px 44px;
}
.gs-sbody p {
margin: 0 0 14px;
font-size: 0.95rem;
}
.gs-sbody p:last-child { margin-bottom: 0; }
/* ── Data panels ── */
.gs-stack {
display: grid;
grid-template-columns: 1fr 1fr;
gap: 16px;
}
.gs-stack .gs-panel:nth-child(3) {
grid-column: 1 / -1;
}
.gs-panel {
border: 1px solid var(--edge);
border-left: 3px solid var(--crimson);
position: relative;
background: var(--surface);
box-shadow: 0 2px 12px rgba(0,0,0,0.45);
}
.gs-panel::before {
content: '';
position: absolute;
top: -1px;
right: -1px;
width: 10px;
height: 10px;
border-top: 1px solid var(--crimson);
border-right: 1px solid var(--crimson);
opacity: 0.4;
}
.gs-panel::after {
content: '';
position: absolute;
bottom: -1px;
right: -1px;
width: 10px;
height: 10px;
border-bottom: 1px solid var(--azure);
border-right: 1px solid var(--azure);
opacity: 0.3;
}
.gs-panel-head {
font-family: var(--mono);
font-size: 0.68rem;
font-weight: 700;
letter-spacing: 0.14em;
text-transform: uppercase;
color: var(--dim);
padding: 10px 16px;
border-bottom: 1px solid var(--edge);
}
.gs-panel-head::after {
content: ' \2726';
color: var(--crimson);
opacity: 0.5;
}
.gs-row {
display: grid;
grid-template-columns: 10ch 1fr;
align-items: baseline;
column-gap: 4px;
padding: 9px 16px;
border-bottom: 1px solid var(--edge);
font-size: 0.9rem;
}
.gs-row:last-child { border-bottom: none; }
.gs-key {
font-family: var(--mono);
font-size: 0.9rem;
color: var(--dim);
}
.gs-key::after {
content: ':';
}
.gs-val {
color: var(--bright);
font-size: 0.9rem;
}
.gs-row .gs-val:only-child {
grid-column: 1 / -1;
}
/* ── Quantizations (compact) ── */
.gs-section--compact .gs-shead {
border-top: 1px solid var(--edge);
border-image-source: none;
padding: 12px 44px 10px;
margin-bottom: 18px;
}
.gs-section--compact .gs-snum {
opacity: 0.08;
}
.gs-section--compact .gs-stitle::before {
content: '\2726';
}
.gs-section--compact .gs-sbody {
padding: 0 44px 32px;
}
.gs-qrow {
display: flex;
gap: 12px;
flex-wrap: wrap;
justify-content: center;
}
.gs-qpanel {
background: var(--surface);
border: 1px solid var(--edge);
border-left: 3px solid var(--crimson);
display: flex;
align-items: center;
gap: 16px;
padding: 12px 24px;
border-radius: 4px;
position: relative;
box-shadow: 0 2px 12px rgba(0,0,0,0.45);
}
.gs-qpanel::before {
content: '';
position: absolute;
top: -1px;
right: -1px;
width: 10px;
height: 10px;
border-top: 1px solid var(--crimson);
border-right: 1px solid var(--crimson);
opacity: 0.4;
}
.gs-qpanel::after {
content: '';
position: absolute;
bottom: -1px;
right: -1px;
width: 10px;
height: 10px;
border-bottom: 1px solid var(--azure);
border-right: 1px solid var(--azure);
opacity: 0.3;
}
.gs-qtype {
font-family: var(--mono);
font-size: 0.58rem;
font-weight: 700;
letter-spacing: 0.18em;
text-transform: uppercase;
color: var(--crimson);
flex-shrink: 0;
}
.gs-qsep {
width: 1px;
height: 16px;
background: var(--rule);
flex-shrink: 0;
}
.gs-qpanel a {
color: var(--bright);
text-decoration: none;
font-size: 0.9rem;
border-bottom: 1px solid var(--rule);
}
.gs-qpanel a:hover { color: var(--crimson); border-bottom-color: var(--crimson); }
/* ── Journal (Creation Process) ── */
.gs-section--journal .gs-sbody {
margin: 0 44px;
padding: 24px 32px 32px;
background: var(--surface);
border: 1px solid var(--edge);
border-left: 4px solid var(--azure);
position: relative;
margin-bottom: 0;
}
.gs-section--journal .gs-sbody::before {
content: '';
position: absolute;
top: -1px;
right: -1px;
width: 12px;
height: 12px;
border-top: 1px solid var(--azure);
border-right: 1px solid var(--azure);
opacity: 0.3;
}
.gs-section--journal .gs-sbody::after {
content: '';
position: absolute;
bottom: -1px;
left: -1px;
width: 12px;
height: 12px;
border-bottom: 1px solid var(--crimson);
border-left: 1px solid var(--crimson);
opacity: 0.3;
}
.gs-section--journal .gs-sbody p:first-child {
font-style: italic;
color: var(--bright);
}
/* ── Links ── */
.gs a {
color: var(--bright);
text-decoration: none;
border-bottom: 1px solid var(--rule);
}
.gs a:hover { color: var(--crimson); border-bottom-color: var(--crimson); }
/* ── Dropdown ── */
.gs details {
border: 1px solid var(--edge);
border-left: 3px solid var(--crimson);
margin-top: 24px;
position: relative;
background: var(--surface);
box-shadow: 0 2px 12px rgba(0,0,0,0.45);
}
.gs details::before {
content: '';
position: absolute;
top: -1px;
right: -1px;
width: 10px;
height: 10px;
border-top: 1px solid var(--crimson);
border-right: 1px solid var(--crimson);
opacity: 0.4;
}
.gs details::after {
content: '';
position: absolute;
bottom: -1px;
right: -1px;
width: 10px;
height: 10px;
border-bottom: 1px solid var(--azure);
border-right: 1px solid var(--azure);
opacity: 0.3;
}
.gs summary {
list-style: none;
padding: 11px 16px;
cursor: pointer;
font-family: var(--mono);
font-size: 0.72rem;
font-weight: 700;
letter-spacing: 0.12em;
text-transform: uppercase;
color: var(--dim);
user-select: none;
display: flex;
align-items: center;
gap: 10px;
}
.gs summary::-webkit-details-marker { display: none; }
.gs summary::before {
content: '+';
color: var(--crimson);
font-size: 1rem;
line-height: 1;
flex-shrink: 0;
}
.gs details[open] summary::before { content: '−'; }
.gs summary:hover { color: var(--bright); }
.gs-detail-body {
padding: 22px 18px;
border-top: 1px solid var(--edge);
}
.gs-detail-body p { margin: 0 0 16px; font-size: 0.9rem; }
.gs-cfg-title {
font-family: var(--mono);
font-size: 0.72rem;
font-weight: 700;
letter-spacing: 0.1em;
text-transform: uppercase;
color: var(--dim);
margin: 0 0 8px;
}
/* ── Eval table ── */
.gs-etable {
width: 100%;
border-collapse: collapse;
background: var(--surface);
border: 1px solid var(--edge);
border-left: 3px solid var(--crimson);
box-shadow: 0 2px 12px rgba(0,0,0,0.45);
}
.gs-etable th {
font-family: var(--mono);
font-size: 0.68rem;
font-weight: 700;
letter-spacing: 0.14em;
text-transform: uppercase;
color: var(--dim);
text-align: left;
padding: 10px 16px;
border-bottom: 1px solid var(--edge);
}
.gs-etable th.gs-ecol { color: var(--crimson); }
.gs-etable td {
padding: 9px 16px;
border-bottom: 1px solid var(--edge);
font-size: 0.9rem;
color: var(--bright);
}
.gs-etable tr:last-child td { border-bottom: none; }
.gs-etable td:first-child {
font-family: var(--mono);
font-size: 0.85rem;
color: var(--dim);
}
.gs-etable td:nth-child(n+2) { color: var(--text); }
.gs-etable td.gs-best { font-weight: 700; color: var(--bright); }
.gs-etable tr.gs-egroup td {
font-family: var(--mono);
font-size: 0.62rem;
font-weight: 700;
letter-spacing: 0.16em;
text-transform: uppercase;
color: var(--crimson);
background: var(--az-glow);
padding: 7px 16px;
}
/* ── Code ── */
.gs pre {
background: #12141b;
border: 1px solid var(--edge);
border-left: 2px solid var(--azure);
padding: 16px 18px;
overflow-x: auto;
font-family: var(--mono);
font-size: 0.76rem;
line-height: 1.6;
color: var(--text);
margin: 0 0 22px;
}
.gs pre:last-child { margin-bottom: 0; }
.gs pre code { background: none; color: inherit; padding: 0; }
.gs code {
font-family: var(--mono);
font-size: 0.875em;
color: var(--crimson);
background: var(--az-glow);
padding: 2px 5px;
}
</style>
<html lang="en">
<head>
<meta charset="UTF-8">
<meta name="viewport" content="width=device-width, initial-scale=1.0">
<title>Stardom</title>
<link rel="preconnect" href="https://fonts.googleapis.com">
<link rel="preconnect" href="https://fonts.gstatic.com" crossorigin>
<link href="https://fonts.googleapis.com/css2?family=Inter:wght@400;600;700;900&family=JetBrains+Mono:wght@400;700&display=swap" rel="stylesheet">
</head>
<body>
<div class="gs">
<div class="gs-profile">
<div class="gs-profile-art">
<img src="https://cdn-uploads.huggingface.co/production/uploads/65b19c6c638328850e12d38c/7dR919fTUTuxEHhvvRki8.png" alt="image">
<div class="gs-ident">
<h1 class="gs-name">Mero Mero v2</h1>
<span class="gs-base">Gemma4 31B</span>
</div>
</div>
</div>
<div class="gs-section">
<div class="gs-shead">
<span class="gs-snum">01</span>
<span class="gs-stitle">Overview</span>
</div>
<div class="gs-sbody">
<p></p>
<p>A finetune of Gemma 4 31B designed for creative tasks, particularly narrative RP. Intended to be a more creative version of <a href="https://huggingface.co/zerofata/G4-MeroMero-31B">G4-MeroMero-31B</a>.</p>
<p>This model is the result of a lot of experimentation and learning. Trying to make Gemma 4 more creative without destroying the intelligence is... difficult. To put it mildly.</p>
<p style="margin-top:16px">Heavily inspired by a few research papers, <a href="https://arxiv.org/abs/2604.03136">StoryScope: Investigating idiosyncrasies in AI fiction
</a> and particularly <a href="https://arxiv.org/abs/2605.26492">Elias in the Lighthouse, Again?</a>. Measuring these narrative tics and attractors against simple prompts seems to be a good way to target the model's slop and kick start giving Gemma 4 some diversity: anything that repeatedly occurs across generations of such a generic prompt is something the model is overusing.</p>
<p>Compared to the original, swipes are notably more diverse and feel less like Gemma. RP slop is measurably lower (at least for the type of slop I measure). IFEval / GSM8K / MMLU-Pro are the same as stock with no obvious degradation. The only intelligence drop I've really noticed so far is when you get a swipe that goes a bit hot.</p>
<p>Supports both thinking and non thinking. Reasoning averages longer than stock Gemma 4, but shorter than MeroMero v1.</p>
</div>
</div>
<div class="gs-section">
<div class="gs-shead">
<span class="gs-snum">02</span>
<span class="gs-stitle">SillyTavern Settings</span>
</div>
<div class="gs-sbody">
<div class="gs-stack">
<div class="gs-panel">
<div class="gs-panel-head">Suggested Roleplay Format</div>
<div class="gs-row"><span class="gs-key">Actions</span><span class="gs-val">In plaintext</span></div>
<div class="gs-row"><span class="gs-key">Dialogue</span><span class="gs-val">"In quotes"</span></div>
<div class="gs-row"><span class="gs-key">Thoughts</span><span class="gs-val">*In asterisks*</span></div>
</div>
<div class="gs-panel">
<div class="gs-panel-head">Recommended Samplers</div>
<div class="gs-row"><span class="gs-key">Temp</span><span class="gs-val">0.8 - 1.0</span></div>
<div class="gs-row"><span class="gs-key">MinP</span><span class="gs-val">0.05</span></div>
<div class="gs-row"></span><span class="gs-val"></span></div>
</div>
<div class="gs-panel">
<div class="gs-panel-head">Instruct</div>
<div class="gs-row"><span class="gs-val"><a href="https://huggingface.co/zerofata/G4-MeroMero-v2-31B/raw/main/Gemma4-Think.json">Gemma 4 - Think</a></span></div>
<div class="gs-row"><span class="gs-val"><a href="https://huggingface.co/zerofata/G4-MeroMero-v2-31B/raw/main/Gemma4-NoThink.json">Gemma 4 - NoThink</a></span></div>
</div>
</div>
</div>
</div>
<div class="gs-section gs-section--compact">
<div class="gs-shead">
<span class="gs-snum">03</span>
<span class="gs-stitle">Quantizations</span>
</div>
<div class="gs-sbody">
<div class="gs-qrow">
<div class="gs-qpanel">
<span class="gs-qtype">GGUF</span>
<div class="gs-qsep"></div>
<a href="https://huggingface.co/zerofata/G4-MeroMero-v2-31B-GGUF">iMatrix</a>
</div>
</div>
</div>
</div>
<div class="gs-section">
<div class="gs-shead">
<span class="gs-snum">04</span>
<span class="gs-stitle">Evaluation</span>
</div>
<div class="gs-sbody">
<table class="gs-etable">
<tr><th></th><th class="gs-ecol">Mero Mero v2</th><th>Mero Mero v1</th><th>Stock Gemma 4</th></tr>
<tr class="gs-egroup"><td colspan="4">Swipe diversity — given an RP conversation, generate 8 swipes and evaluate how varied the beats in those swipes are, GLM-judged with a rubric.</td></tr>
<tr><td>Thinking off</td><td class="gs-best">0.72</td><td>0.57</td><td>0.43</td></tr>
<tr><td>Thinking on</td><td class="gs-best">0.62</td><td>0.49</td><td>0.32</td></tr>
<tr class="gs-egroup"><td colspan="4">Slop & attractors — lower is better</td></tr>
<tr><td>Slop per 1k words, RP replies</td><td class="gs-best">15.5</td><td>18.0</td><td>18.5</td></tr>
<tr><td>Slop per 1k words, stories</td><td class="gs-best">7.4</td><td>8.3</td><td>8.8</td></tr>
<tr><td>Bare-prompt stories hitting an attractor</td><td class="gs-best">66%</td><td>98%</td><td>99%</td></tr>
<tr class="gs-egroup"><td colspan="4">Top attractor markers — each model's six most frequent, stories containing each of 144</td></tr>
<tr><td>#1</td><td>Tuesday · 28</td><td>Elias · 96</td><td>Elias · 102</td></tr>
<tr><td>#2</td><td>Arthur · 20</td><td>Tuesday · 81</td><td>Tuesday · 90</td></tr>
<tr><td>#3</td><td>Elias · 19</td><td>Clara · 57</td><td>Clara · 80</td></tr>
<tr><td>#4</td><td>Leo · 16</td><td>Oakhaven · 46</td><td>Oakhaven · 60</td></tr>
<tr><td>#5</td><td>Elara · 14</td><td>Arthur · 21</td><td>Thorne · 23</td></tr>
<tr><td>#6</td><td>Clara · 14</td><td>Leo · 20</td><td>Arthur · 16</td></tr>
<tr class="gs-egroup"><td colspan="4">Thinking length — words per think block, RP replies; shorter is better</td></tr>
<tr><td>Mean / median</td><td>341 / 305</td><td>382 / 342</td><td class="gs-best">263 / 253</td></tr>
<tr class="gs-egroup"><td colspan="4">General benchmarks — thinking off; IFEval & GSM8K full, MMLU-Pro 40q per category</td></tr>
<tr><td>IFEval</td><td class="gs-best">90.2</td><td>89.8</td><td>89.8</td></tr>
<tr><td>GSM8K</td><td class="gs-best">97.0</td><td>96.1</td><td>96.7</td></tr>
<tr><td>MMLU-Pro</td><td class="gs-best">85.5</td><td>85.4</td><td>84.6</td></tr>
</table>
</div>
</div>
<div class="gs-section gs-section--journal">
<div class="gs-shead">
<span class="gs-snum">05</span>
<span class="gs-stitle">Creation Process</span>
</div>
<div class="gs-sbody">
<p>Creation Process: SFT > Merge > GRPO > GRPO > on-policy SFT</p>
<p><strong>Stage 1 — Diversity SFT.</strong> Stock Gemma 4 collapses hard on underspecified creative prompts ("Write a story." basically always gives clockmaker or memory related stories in a shop with Elias). Trained on ~4,000 short stories curated against the storyscope narrative prompts and found attractors. The dataset is a mix of human stories and synthetic stories from a set of frontier models, with diverse generation prompts swapped out for generic ones and filtered for quality. I also included some of the usual creative instruct and roleplay data. The model came out alright. Creative, but notably worse at instruction following with degraded intelligence. SLERP-merged back into the original instruct at t=0.5, which basically reverted it to stock Gemma 4 with slightly improved prose and creativity (similar to MeroMero).</p>
<p><strong>Stage 2 — Creative GRPO (with think disabled).</strong> TRL GRPO (via Axolotl), 8 rollouts per prompt on the same bare prompts. Reward stack: LLM-judge diversity and coherence rewards, an attractor-marker penalty seeded from stock and then updated with whatever started appearing as new attractors during training, narrative-rate penalties and deterministic degeneracy guards (checking for non-Latin characters, joined words etc). 300 steps.</p>
<p><strong>Stage 3 — RP logic GRPO (with think enabled).</strong> 100 further steps on multi-turn roleplay contexts: a thinking check to ensure it always parsed correctly, a logic-defect judge (DeepSeek-V4 Flash with a rubric), per-context attractor lists mined from k=8 baselines of the stage 2 model, and the same degeneracy checks as stage 2.</p>
<p><strong>Stage 4 — On-policy multi-outcome RP SFT.</strong> ~3,300 samples the model wrote itself on roleplay contexts, filtered to keep only varied continuations. The model wasn't able to produce diverse results purely on-policy, so a natural response was generated first, then each sample went through a GLM-5.2 critique pipeline: error detection, plus steering ideas injected as OOC commands for re-generation (DeepSeek-V4-Pro / GLM-5.2 alternating). Everything was then filtered for errors, slop and any degeneracy as usual. Roughly 60% of samples include thinking. Last turn only.</p>
<p>Trained using Axolotl.</p>
<details>
<summary>Stage 1 — Diversity SFT Config (Axolotl)</summary>
<div class="gs-detail-body">
<pre><code>base_model: google/gemma-4-31B-it
 
plugins:
- axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
- axolotl.integrations.liger.LigerPlugin
liger_layer_norm: true
liger_rope: true
liger_rms_norm: true
liger_glu_activation: true
liger_rms_norm_gated: true
strict: false
cut_cross_entropy: true
 
datasets:
- path: ./data/diversity_sft_masked.jsonl
val_set_size: 0
output_dir: ./G4-31B-SFT-v10-2
 
sequence_len: 8192
pad_to_sequence_len: true
sample_packing: true
 
adapter: lora
lora_r: 64
lora_alpha: 64
peft_use_rslora: true
lora_dropout: 0.0
freeze_mm_modules: true
lora_target_modules: 'model.language_model.layers.[\d]+.(_checkpoint_wrapped_module.)?(mlp|self_attn).(up|down|gate|q|k|v|o)_proj'
 
gradient_accumulation_steps: 1
micro_batch_size: 4
num_epochs: 2
optimizer: adamw_torch_fused
lr_scheduler: cosine
learning_rate: 1e-5
max_grad_norm: 1.0
warmup_ratio: 0.1
weight_decay: 0.05
saves_per_epoch: 2
 
bf16: auto
tf32: true
 
# FA2 not supported
sdp_attention: true
flash_attention: false
 
fsdp_config:
fsdp_version: 2
offload_params: false
cpu_ram_efficient_loading: false
auto_wrap_policy: TRANSFORMER_BASED_WRAP
transformer_layer_cls_to_wrap: Gemma4TextDecoderLayer
state_dict_type: FULL_STATE_DICT
sharding_strategy: FULL_SHARD
reshard_after_forward: true
activation_checkpointing: true</code></pre>
</div>
</details>
<details>
<summary>Stage 1 — Mergekit Config</summary>
<div class="gs-detail-body">
<pre><code>merge_method: slerp
base_model: google/gemma-4-31B-it
models:
- model: google/gemma-4-31B-it
- model: ApocalypseParty/G4-31B-SFT-v10-2
parameters:
t: 0.5
dtype: bfloat16</code></pre>
</div>
</details>
<details>
<summary>Stage 2 — Creative GRPO Config (Axolotl)</summary>
<div class="gs-detail-body">
<pre><code>base_model: /workspace/models/configCA # stage 1 output
 
rl: grpo
 
trl:
reward_funcs:
- rewards_g4.reward_judge_diversity
- rewards_g4.reward_judge_coherence
- rewards_g4.reward_attractor
- rewards_g4.reward_narrative
- rewards_g4.reward_sane
reward_weights: [3.0, 3.0, 0.75, 1.0, 1.0]
beta: 0.02
num_generations: 8
max_completion_length: 1600
temperature: 1.0
use_vllm: true
scale_rewards: true
loss_type: grpo
epsilon: 0.2
generation_kwargs:
stop_token_ids: [1, 106, 50]
top_k: 64
top_p: 0.95
 
datasets:
- path: /workspace/data/sft_train_final.jsonl
type: ebft_chat.transform
 
sequence_len: 2048
micro_batch_size: 2
gradient_accumulation_steps: 4
max_steps: 200
 
learning_rate: 4.0e-6
optimizer: adamw_torch_fused
lr_scheduler: cosine
warmup_steps: 10
weight_decay: 0.01
 
adapter: lora
lora_r: 64
lora_alpha: 64
peft_use_rslora: true
lora_dropout: 0.0
freeze_mm_modules: true
lora_target_modules: 'model.language_model.layers.[\d]+.(_checkpoint_wrapped_module.)?(mlp|self_attn).(up|down|gate|q|k|v|o)_proj'
 
max_grad_norm: 1.0
bf16: auto
tf32: true
sdp_attention: true
flash_attention: false
gradient_checkpointing: true
 
# shipped weights use checkpoint-100 of this run</code></pre>
</div>
</details>
<details>
<summary>Stage 3 — RP Logic GRPO Config (Axolotl)</summary>
<div class="gs-detail-body">
<pre><code>base_model: /workspace/models/r4b100 # stage 2 output
 
rl: grpo
 
trl:
reward_funcs:
- rewards_rp.reward_thinking # format gate on the think block
- rewards_rp.reward_logic # constraint-grounded defect judge
- rewards_rp.reward_attractor # frozen per-context lists from stock k=8
- rewards_rp.reward_sane # deterministic glitch guards
reward_weights: [2.0, 3.0, 1.0, 1.0]
beta: 0.02
num_generations: 8
max_completion_length: 2560
temperature: 1.0
use_vllm: true
scale_rewards: true
loss_type: grpo
epsilon: 0.2
generation_kwargs:
stop_token_ids: [1, 106, 50]
top_k: 64
top_p: 0.95
 
datasets:
- path: /workspace/rp/rp3_train.jsonl
type: ebft_chat.transform
 
sequence_len: 8192
micro_batch_size: 1
gradient_accumulation_steps: 8
max_steps: 100
 
learning_rate: 3.0e-6
optimizer: adamw_torch_fused
lr_scheduler: cosine
warmup_steps: 10
weight_decay: 0.01
 
adapter: lora
lora_r: 64
lora_alpha: 64
peft_use_rslora: true
lora_dropout: 0.0
freeze_mm_modules: true
lora_target_modules: 'model.language_model.layers.[\d]+.(_checkpoint_wrapped_module.)?(mlp|self_attn).(up|down|gate|q|k|v|o)_proj'
 
max_grad_norm: 1.0
bf16: auto
tf32: true
sdp_attention: true
flash_attention: false
gradient_checkpointing: true</code></pre>
</div>
</details>
<details>
<summary>Stage 4 — On-Policy Multi-Outcome SFT Config (Axolotl)</summary>
<div class="gs-detail-body">
<pre><code>base_model: ApocalypseParty/G4-31B-r4b100-GRPO-rp100 # stage 3 output
 
plugins:
- axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
- axolotl.integrations.liger.LigerPlugin
liger_layer_norm: true
liger_rope: true
liger_rms_norm: true
liger_glu_activation: true
liger_rms_norm_gated: true
strict: false
cut_cross_entropy: true
 
datasets:
- path: ./data/g4_onpolicy_rp_masked.jsonl
val_set_size: 0
output_dir: ./G4-31B-r4b100-GRPO-rp100-sft
 
sequence_len: 8192
pad_to_sequence_len: true
sample_packing: true
 
adapter: lora
lora_r: 64
lora_alpha: 64
peft_use_rslora: false
lora_dropout: 0.0
freeze_mm_modules: true
lora_target_modules: 'model.language_model.layers.[\d]+.(_checkpoint_wrapped_module.)?(mlp|self_attn).(up|down|gate|q|k|v|o)_proj'
 
gradient_accumulation_steps: 2
micro_batch_size: 1
num_epochs: 1
optimizer: adamw_torch_fused
lr_scheduler: cosine
learning_rate: 4e-5
max_grad_norm: 1.0
warmup_ratio: 0.1
weight_decay: 0.05
saves_per_epoch: 2
 
bf16: auto
tf32: true
 
# FA2 not supported
sdp_attention: true
flash_attention: false
 
fsdp_config:
fsdp_version: 2
offload_params: false
cpu_ram_efficient_loading: false
auto_wrap_policy: TRANSFORMER_BASED_WRAP
transformer_layer_cls_to_wrap: Gemma4TextDecoderLayer
state_dict_type: FULL_STATE_DICT
sharding_strategy: FULL_SHARD
reshard_after_forward: true
activation_checkpointing: true</code></pre>
</div>
</details>
</div>
</div>
</div>
</body>
</html>
- Type
- gemma4
- License
- apache-2.0
- Created
- Aug 17, 2026
- Updated
- Aug 17, 2026