Downloads · 30 days
1.4K
80% of all-time downloads
thestarfarer/Ministral-3-14B-lil-logbook
Ministral-3-14B-lil-logbook is a machine learning model from thestarfarer. Use it for the machine learning task on the model card, and read the license before you ship it in a product. It is set up for peft. The card lists the license as apache-2.0.
e2 beats the base model on 80% of prompts (91 of 114; blind judges, prompt-matched, 5 rolls each).
Downloads · 30 days
1.4K
80% of all-time downloads
All-time downloads
1.8K
Public
Repo size
435 GB
Likes
0
Public
Click a slice to open those files.
.gguf196 GB · 49%
From the Hugging Face model README
e2 beats the base model on 80% of prompts (91 of 114; blind judges, prompt-matched, 5 rolls each).
| model | consistency | coherence | format | craft | best-pick | paired vs p3 |
|---|---|---|---|---|---|---|
| e2 | 1.61 | 1.41 | 1.06 | 1.03 | 32% | +0.31 (t 3.9) |
| e1 | 1.54 | 1.36 | 1.10 | 1.00 | 31% | +0.21 (t 2.7) |
| base | 1.33 | 1.15 | 1.02 | 0.88 | 20% | -0.41 (t -4.1) |
| p3 | 1.57 | 1.34 | 0.93 | 0.96 | 17% | - |
Blinded LLM-judge panel: the long-context chapter's generations plus base.
Grab gguf/Ministral-3-14B-length-e2-Q8_0.gguf (14.4 GB): e2, the program's final model.
Format: raw text completion, no chat template. Every turn is a # x header line (a short speaker code) followed by the post; end the prompt with the bare header of the speaker to write next:
# k
*She drops the satchel on the table.* "You're late."
# m
"Traffic." He shrugs off his coat. "What did I miss?"
# k
Add \n# or just # as a stop string.
Samplers to start with: temperature 0.7, min_p 0.1, repetition penalty 1.07 over 128 tokens, DRY 0.5 / 1.75 / allowed length 2 / range 1024 with sequence breakers ["\n", ":", "\"", "*"]; everything else neutral. Response 350 tokens, context 8192.
LoRA adapters trained on the same dataset, one variable changed at a time:
the three original runs below, a learning-rate ladder, a full rank × LR
grid under lr-sweep/ (24 adapters, named r<rank>-lr<lr>), a three-source
mix (mix/, v3only/), and a schedule palette (pal/). F16/Q8_0 GGUFs
for the original three.
The original three runs were evaluated on loss curves and spot generations only; the ladder and grid add an out-of-domain loss axis and a blinded generation-level composition probe (sections below).
Multi-turn dialogue packed to 512-token samples — 13,667 samples,
6.16M tokens total. Turns are delimited by
# <code> headers — short per-character abbreviations, not names. Unattributed
narration is folded into the speaker's turn:
# m
Okay.
This is as far as I come.
# s
We come to a stop in front of a tree positioned in the middle of three spotlights.
It's not a particularly interesting tree or anything like that, so I'm a little confused as to why it's been so carefully illuminated.
What is this?
# m
A tree, idiot.
# s
No, like...what is this?
The pack was built in two phase-shifted passes — the second pass offsets every chunk boundary by 256 tokens, so the corpus appears twice with different windowing and exact-duplicate chunks removed. Experimental idea for cheap augmentation on small corpora; its actual effect is not isolated or evaluated in this repo.
Shared config: rsLoRA r512 / α512, dropout 0, targets q,k,v,o,gate,up,down,
batch 16 (no accumulation), 838 steps (one pass), constant LR after warmup,
BF16 base + FP32 AdamW, seed 42, eval on 256 held-out samples every 25 steps.
| run | path | init | LR | warmup | max_grad_norm | final eval |
|---|---|---|---|---|---|---|
v3-lil | / | continued from writer-sft-rp-v3 | 7.5e-6 | 30 | 10.0 | 1.530 |
base-lil | /base-lil/ | fresh adapter on base | 5e-6 | 35 | 10.0 | 1.582 |
base-lil-clip1 | /base-lil-clip1/ | fresh adapter on base | 5e-6 | 35 | 1.0 | 1.583 |
Init barely matters — according to metrics. base-lil matched v3-lil's
eval curve point-for-point for the first 250 steps and finished 0.05 behind at
33% lower LR — the 186M-token RP adapter it continues from bought approximately
nothing on this corpus by loss. Loss can't tell the whole story here: when the
two models miss, they might miss differently.
Gradient clipping at 1.0 changed nothing. base-lil-clip1 clipped on 100% of
steps (raw grad norms mean 14.9, max 36) and still landed within 0.001 of the
unclipped control. AdamW divides the first moment by the root second moment, so a
rescale applied to every gradient approximately cancels out of the update — clipping only
matters when it fires intermittently.
lr-sweep/)The question: how low can LR go while the base model still learns this corpus —
smaller weight modification for relatively the same result, because every unit
of fit is paid for in general-capability damage. Fresh adapters on base, same
recipe as base-lil-clip1 (max_grad_norm 1.0, justified above), plus an
eval before step 0 (the untrained baseline all deltas are measured from)
and a second eval set: 256 held-out forum-RP samples from the
v3
corpus, line-truncated to 512 ctx — out-of-domain prose the training never
shows, scored by the same eval machinery. In-domain fit and out-of-domain
damage, side by side:
| LR | in-domain eval | OOD prose eval | gained | prose lost | gained/lost |
|---|---|---|---|---|---|
| — (untrained base) | 2.421 | 2.655 | — | — | — |
r512-lr1e-7 | 2.015 | 2.759 | 0.405 | +0.104 | 3.91 |
r512-lr2e-7 | 1.970 | 2.811 | 0.451 | +0.156 | 2.89 |
r512-lr5e-7 | 1.898 | 2.871 | 0.523 | +0.216 | 2.42 |
r512-lr2e-6 | 1.729 | 2.977 | 0.692 | +0.322 | 2.15 |
5e-6 (= base-lil-clip1) | 1.583 | 3.067 | 0.838 | +0.412 | 2.04 |
No cliff down to 1e-7, and the exchange rate improves monotonically as LR drops. A 50× LR cut keeps 48% of the learning for 25% of the damage. Forgetting here is distribution shift, not training per se: nothing in a 100%-narrow stream defends the out-of-domain directions, so damage scales with total movement — and lower LR simply moves less.
Loss can't answer the question that matters for an RP adapter: do the learned
personas compose with content the corpus never contained? Probe: 32 scenes
placing the corpus's recurring speakers (by their # code headers) into
mundane/technical situations foreign to the corpus — broken spreadsheets,
router channels, tax forms, bed-leveling a 3D printer — each ending on a
character header, each embedding a concrete technical fact the continuation can
get right or wrong. 5 rerolls × 6 checkpoints (base, the four ladder runs,
base-lil-clip1), scored 0–2 on voice (unmistakably that speaker), coherence
(engages the scenario's content correctly), format (clean header structure,
no degeneration) by an LLM-judge panel.
Sampler first, because it dominated round one. The sampler recommended for
v3 (temp 0.5, min_p 0.1, rep 1.03) is a loop incubator on these base-derived
checkpoints: 30–65% of completions collapsed into verbatim loops or ...
header volleys, at every LR. At temp 0.8, min_p 0.05, rep 1.1 the collapse
rate drops to ≤1.2% everywhere. Use the latter with these adapters.
Scoring was only trusted blinded. Per-checkpoint judging (one judge instance per model) swung good-rates by ±8pp from judge strictness alone, so the contested rungs were re-scored as a pooled, shuffled, label-stripped set with each judge seeing a mix of all three. Blinded good-rate (voice+coherence+format ≥ 5/6), t0.8 sampler, n=160 per rung:
lr5e-7 | lr1e-7 | lr2e-7 | |
|---|---|---|---|
| good-rate | 54% | 50% | 43% |
| mean score (0–6) | 4.53 | 4.38 | 4.28 |
Differences are ≤2σ at this n. Read: 5e-7 is the best point estimate; 1e-7 matches it within noise at half the prose damage. Bare base scores 7% — it cannot hold the header format at all, leaking pretraining documents mid-scene, so some adapter is load-bearing for the format itself. The universal bottleneck at every LR is coherence (~1.0/2): completions open on-task, then drift back into the corpus's relationship register within a few turns. That is the same distribution-shift mechanism as the prose-loss column, observed at generation level — and the motivation for the replay-mixing item below.
lr-sweep/r*-lr*)Rank 64 / 128 / 256 / 512 / 1024 crossed with LR 1e-7 / 2e-7 / 5e-7 / 2e-6 / 5e-6,
same recipe as the ladder. Alpha is scale-matched, α = √(512·r)
(r64→181, r128→256, r256→362, r512→512, r1024→724): with rsLoRA the forward gain
is α/√r, so this holds it at 22.6 for every rank and the axis measures capacity
alone, not capacity times step size. r1024 does not fit 80 GB with fp32 AdamW and
uses paged_adamw_32bit (same math, CPU-pageable state). The r512 row was
trained on an H100 and the rest on an A100 with a newer torch; their untrained
baselines differ by ~0.005, so cross-row comparisons carry that offset.
Loss: in-domain gained / OOD prose lost (from the untrained baselines):
| rank | 1e-7 | 2e-7 | 5e-7 | 2e-6 | 5e-6 |
|---|---|---|---|---|---|
| r64 | 0.285 / +0.002 | 0.337 / +0.046 | 0.390 / +0.104 | 0.482 / +0.202 | 0.555 / +0.254 |
| r128 | 0.331 / +0.035 | 0.370 / +0.074 | 0.428 / +0.135 | 0.532 / +0.228 | 0.635 / +0.292 |
| r256 | 0.366 / +0.067 | 0.408 / +0.114 | 0.471 / +0.176 | 0.601 / +0.271 | 0.730 / +0.340 |
| r512 | 0.405 / +0.104 | 0.451 / +0.156 | 0.523 / +0.216 | 0.692 / +0.322 | 0.838 / +0.412 |
| r1024 | 0.451 / +0.151 | 0.502 / +0.197 | 0.591 / +0.264 | 0.793 / +0.393 | 0.882 / +0.590 |
(r512 @ 5e-6 is base-lil-clip1.) Both axes are monotone in every row and
column. Rank and LR act like one budget of total weight movement: cells along
the same diagonal land on nearly the same trade (r64 @ 5e-7 ≈ r512 @ 1e-7).
The gentle corner learns for free: r64 @ 1e-7 gains 0.285 for +0.002. Its OOD
prose loss even dips below baseline for the first ~300 steps before giving
it back, so early stopping there is positive-sum.
Composition: good-rate, mundane / fiction scenes (%). The probe was widened to 64 scenes: the 32 mundane/technical ones plus 32 that place the same speakers in other fictional settings (dungeon, starship, noir, heist, period pieces...), each with a checkable in-world fact. 5 rerolls, t0.8 sampler, all cells judged in one pooled, shuffled, label-stripped panel, n=160 per cell per scene set:
| rank | 1e-7 | 2e-7 | 5e-7 | 2e-6 | 5e-6 |
|---|---|---|---|---|---|
| r64 | 37 / 33 | 44 / 48 | 49 / 54 | 48 / 54 | 51 / 49 |
| r128 | 39 / 37 | 48 / 49 | 56 / 55 | 50 / 48 | 47 / 39 |
| r256 | 47 / 44 | 49 / 49 | 50 / 53 | 55 / 51 | 45 / 37 |
| r512 | — | — | 57 / 48 | — | — |
| r1024 | 50 / 59 | 52 / 52 | 45 / 40 | 18 / 16 | 16 / 14 |
Untrained base: 9 / 8. The best cells form a diagonal plateau at moderate total movement (r128 @ 5e-7, r256 @ 2e-6, r512 @ 5e-7, r1024 @ 2e-7), tied within noise (SE ≈ 4 pp). The best-fit cells are the worst roleplayers: r1024 @ 2e-6 and 5e-6 hold the lowest in-domain loss in the repo and fall to 14–18%, barely above base. Past the budget optimum, loss and usability anti-correlate. The largest adapter at the gentlest LR, r1024 @ 1e-7, tops the fiction scenes: wide-but-gentle updates install personas that transfer.
When it fries. Format collapse is rare for every trained cell (0–6%); the failure is derailing: coherence scored 0 (ignores, contradicts or abandons the scene). Derail rate:
| rank | 1e-7 | 2e-7 | 5e-7 | 2e-6 | 5e-6 |
|---|---|---|---|---|---|
| r64 | 19% | 18% | 15% | 16% | 18% |
| r128 | 19% | 14% | 16% | 17% | 22% |
| r256 | 17% | 17% | 12% | 16% | 26% |
| r512 | — | — | 17% | — | — |
| r1024 | 13% | 17% | 25% | 46% | 52% |
Base derails 45%. Most of the grid sits on a 12–19% floor. The rate starts climbing when OOD prose damage passes about +0.25 and roughly doubles past +0.35, whatever rank and LR produced the damage. The prose eval is a usable early-warning gauge for frying.
Above 5e-7 a contraction-corruption artifact appears, LR-dependent: 0 corrupted
contractions at ≤5e-7, 14 at 2e-6, 97 at 5e-6, counted across 1,920 probe
completions. Shape: the model emits a contraction stem then neither apostrophe
form — don + corpus-boosted junk (speaker-name fragments, the corpus's
stutter idiom) where 't should be.
Mechanism: the source corpus's own typography runs 10:1 curly (’) to ASCII
(') apostrophes, and the tokenizer makes the two forms competing single
tokens at the same slot ([' don'] + ['’t'|'\'t']). High-LR training suppresses
the ASCII suffix (the curly share of generated apostrophes climbs 25% → 67% up
the ladder), while an ASCII-styled prompt pushes the copy direction the other
way. Usually one side wins; occasionally the conflict flattens the slot and
whatever else the adapter boosted leaks through. The cleanest microscope in
this repo on what the prose-loss column physically is. Fix for a dataset v2:
normalize apostrophes to one form at build time — token counts are unchanged
and the conflict never starts.
mix/, v3only/)Dataset v2, built on the grid's conclusions. The corpus was rebuilt with single-letter randomized turn headers — letters drawn per sample, so persona is untied from header identity and must be carried by the text — and normalized to pure ASCII (curly quotes/apostrophes/dashes folded, 91 remaining non-ASCII samples dropped; this closes the apostrophe-pathology fix): 13,576 samples / 6.12M tokens. Two sources were added in the same 512-token format: forum RP with the same 1-letter headers and a trailing next-speaker header (26,275 / 12.37M) and book prose windows with no headers (5,159 / 2.64M — exactly 1/8 of total). Every batch of 16 is stratified — corpus 4–5, forum 9–10, prose 1–2 — by Bresenham interleave over an ordered file with a sequential sampler, so each step sees the global ratio.
A base-model delimiter probe (5 header schemes × 411 samples, body NLL +
turn-boundary probability) found the # x scheme already near-best; format kept.
| run | steps | LR | corpus eval | forum eval | prose eval |
|---|---|---|---|---|---|
mix/mix-r128-lr5e-7-w100 | 2,797 | 5e-7 const | 2.416 → 2.046 | 2.588 → 2.402 | 2.651 → 2.607 |
v3only/v3only-r128-lr2e-6-w35 | 800 | 2e-6 const | — | 2.588 → 2.397 | 2.661 → 2.614* |
v3only/v3only-r128-lr5e-6-w35 | 800 | 5e-6 const | — | 2.588 → 2.398 | 2.661 → 2.626* |
*v3only rows use the earlier OOD prose eval set; the mix row uses a book-prose eval carved from unused sources. Same machinery, different baselines.
Replay works. The mix run is the first where nothing got worse: book prose improved while both dialogue corpora were being learned. Every cell of the grid had paid prose damage for its learning; mixing 1/8 prose into the stream bought the learning without the bill.
Broad data tolerates LR the narrow corpus couldn't. Forum-only at 2e-6 reached the mix's forum loss in 800 steps instead of 2,797 — and improved its prose eval — where corpus-only training at the same rank and LR had cost +0.23 of prose damage in the grid. 5e-6 bought nothing over 2e-6 step-for-step: the forum corpus is data-limited past 2e-6, not LR-limited.
Loss and preference diverge again. A blinded LLM-judge panel (32 held-out forum-RP prompts × 4 models, shuffled labels, 0–2 rubric on consistency/coherence/format): the mix ranked first — avg rank 1.78, first place on 15/32 — over both forum-only runs (2.38 / 2.56) despite its slightly worse forum loss; the base model (3.28) cannot hold the format at all. The judges' universal complaint across all trained models: speakers blending into each other mid-scene.
5e-7 constant is too low for this breadth. The first quarter of the mix run bought 82% of its corpus learning and 95% of its forum learning; the remaining 2,100 steps bought the rest.
pal/)Four 450-step probes on the first 7,200 rows of the ordered mix (stratification
intact): the constant anchor against three cosine shapes (cosine_with_min_lr,
warmup 30), r128.
| schedule | corpus eval | forum eval | prose eval | panel avg rank |
|---|---|---|---|---|
pal-r128-lr5e-7-w30 (const) | 2.134 | 2.427 | 2.616 | 2.41 |
pal-r128-lr1e-6-w30-cos3e-7 | 2.123 | 2.425 | 2.614 | 2.62 |
pal-r128-lr2e-6-w30-cos6e-7 | 2.099 | 2.420 | 2.616 | 2.69 |
pal-r128-lr1e-6-w30-cos1e-7 | 2.130 | 2.426 | 2.615 | 2.28 |
At this length, the peak is what registers on loss. The deep anneal (1e-6→1e-7) lands on top of the constant anchor on every eval; the aggressive peak (2e-6→6e-7, 2.6× the anchor's total LR) buys the most loss — at zero prose damage, since 450 steps is inside the safe budget for every shape tested. Unconfirmed as a general schedule verdict: 450 steps may simply be too short for the shape of the decay to matter — a long run, where the swing has time to act on the weights, is the real test (see to-dos).
And none of it moves generation quality. The same blinded panel on the four adapters: every rank difference within noise at n=32, and the loss winner placed last on consistency and coherence while holding perfect format. At this scale the schedule is not a lever the judges can feel — the data axis is where generation quality has moved every time.
v3full)The three-source mix above trained on whatever the recut produced. This chapter replaces that bet with a judged one: rebuild the forum corpus at full length, grade every training window individually, and train only on what passes.
Rebuild. The full scrape of the eight forum sources yields 48,351 conversations -> 177,029 recut windows at the 512 cap (6.7x the earlier v3 slice), same cutter rules as before: turn-boundary packing, >=5 body turns + next-speaker header, per-window re-rolled speaker letters.
Judging. A frontier commercial model read each window (25 per call)
and returned a 7-axis verdict: own_side, pure_rp, speaker_names, prose_quality (0-7), coherent, english, questionable. It covered 122,350
windows (69%) before quota economics ended the campaign; the remaining 31%
was judged by the distilled local judge described below, provenance
recorded per row. Pass = all four hard flags clean and pq >= 3
(questionable deliberately ignored - it flags theme, not craft).
Pass rate 39.7% on the teacher's region, 48.5% on the student's (richer sources landed later in the judging order; the student's aggregate matched the teacher's adjacent-region 48.2%, which doubles as a free calibration check). Grade distribution is bottom-heavy: ~44% of windows grade 2 ("functional but flat"), ~35% grade 3, ~11% grade 4, ~2.5% grade 5+.
mixv2/)Design. One ordered file, grown in three increments as judging progressed; a sequential sampler plus checkpoint auto-resume makes the increments one logical run (verified byte-identical against an unbroken run on a 30-step probe before committing). Each batch of 16 is Bresenham-stratified: ~13.3 judged-forum windows, 2 plain-prose replay windows, ~0.7 target-style windows - global ratios hold in every batch.
Config. r128/a256 as before, lr 5e-7 constant, warmup 100, clip 1.0, one pass, 5,259 steps total across the three increments.
Increment 3 filter. By the third increment the distilled judge
supplied token-level probabilities alongside each verdict, so the pass
filter tightened: an argmax-3 whose probability mass hesitated toward 2
(P(pq>=3) < 0.70) was excluded - 3,424 windows evicted. The panels below
suggest this is where the phase's floor gains came from.
| eval set | base | end of inc 1 | inc 2 | inc 3 |
|---|---|---|---|---|
| target-style | 2.415 | 2.193 | 2.161 | 2.135 |
| forum in-dist | 2.588 | 2.418 | 2.402 | 2.398 |
| book prose | 2.651 | 2.607 | 2.598 | 2.600 |
| forum @1k | 2.674 | 2.538 | 2.534 | 2.536 |
| forum @2k | 2.682 | 2.586 | 2.584 | 2.589 |
| forum @4k | 2.680 | 2.603 | 2.603 | 2.610 |
Three findings. (1) Long-context transfer from 512-window training is a one-shot, early-training phenomenon: the 1k/2k/4k evals took nearly all their improvement in increment 1 and flat-lined after - by increment 3 all three tick faintly upward, the data's way of saying the short-window era is over. (2) The target-style eval never plateaued in 5,259 steps. (3) The judged mix reaches 2.398 on the in-distribution eval - past the unfiltered three-source mix's 2.402 - while training on 39-48% of the corpus and discarding the rest. Filtering pays its own bill in loss space; what it buys beyond that is below.
Loss can't see what a user swipes away, so every milestone gets the same blind protocol: 32 held-out prompts per context bucket (512/1024/2048/4096), 5 rerolls each at fixed seed, bf16 base + adapter under vLLM (never the GGUFs). Candidates from the two models under comparison are pooled - ten per prompt - shuffled under fresh letters, and read by a 16-judge blinded LLM panel (one judge per 8-prompt slice) scoring consistency/coherence/format (0-2), naming one best, and listing unusable rolls ("a user would certainly reroll this"). Judge severity drifts between panels, so comparisons are read within-round only; the cross-generation picture is chained from within-round ratios and quoted conservatively.
| 512 | 1024 | 2048 | 4096 | |
|---|---|---|---|---|
| unusable, base (its round) | 27.5% | 35.4% | 20.0% | 16.1% |
| unusable, inc-3 (its round) | 6.2% | 10.8% | 10.0% | 9.3% |
| chained cut vs base | ~4-8x | ~3-4x | ~1.5x | ~1x |
| chained consistency gain | +0.27 | +0.50 | +0.14 | +0.18 |
The shape is a wedge: large at the trained depths, tapering to noise at 4096 - the same shape the eval-loss ladder draws. Each increment bought a different thing: inc 1 the big format/consistency leap off base, inc 2 the short-context floor, inc 3 (the dial-filtered one) floor gains at every depth with the era's best consistency round. None of them moved the ceiling - best-pick counts stayed even in both adapter-vs-adapter rounds - and against the unfiltered three-source mix the judged line now matches its floor (6.2% vs 6.5% unusable at 512, separate rounds) while keeping the markedly higher best-pick share measured head-to-head earlier. Reliability came from data quality; brilliance will have to come from somewhere else.
The dominant failure mode at every depth, in every round, per every judge: breaking established facts - who is where, whose trait is whose, what the scene already ruled out. Format collapse is essentially solved; long-range consistency is the open frontier, which motivates the continuation below.
length/)The corpus re-cut at 1024 and 2048 caps yields 162,620 and 93,837 windows
The plan above was compressed in execution: instead of separate 1024 and
2048 diet phases, the program went straight to the closing wave. e1:
warm start from the increment-3 weights (weights only, fresh optimizer
across the batch-shape change), r128 @ 5e-7 constant, warmup 50, 1,070
steps over one packed file - 14,590 dial-selected corpus windows from
all three context lengths (clean 4+ spine, gap filled by the top of the
3s by P(>=4)), 2,068 matching-length prose-replay rows, and a small
auxiliary RP set - sample-packed to the 2048 cap with block-diagonal
attention masks and per-sample position resets (verified by the
packed-vs-alone per-sample loss test). e2: a true second epoch of
the same file, resumed. Adapters: length/length-r128-lr5e-7-w50 (e1),
length/length-e2-r128-lr5e-7-w50 (e2), resumable checkpoints beside
them.
Losses over the judged-tier battery (q4 = top-graded held-out windows, q2 = low-graded, at each cap; each tier set is only 10 windows, so read these as directional):
| set | p3 (warm start) | e1 | e2 |
|---|---|---|---|
| q4 @ 512 | 2.162 | 2.104 | 2.095 |
| q4 @ 1024 | 2.435 | 2.411 | 2.409 |
| q4 @ 2048 | 2.363 | 2.348 | 2.347 |
| q2 @ 512 | 2.452 | 2.441 | 2.441 |
| q2 @ 1024 | 2.424 | 2.416 | 2.418 |
| q2 @ 2048 | 2.455 | 2.447 | 2.448 |
| in-domain @ 512 | 2.173 | 2.199 | 2.213 |
The q4 sets move 2-6x further than the q2 sets, directionally what dial-selection should produce, though at n=10 per set the margins are thin; the price is a few hundredths on the 512 in-domain set (+0.040 by e2). e1 -> e2 is nearly invisible in loss.
The panel saw more than the loss did. A blinded LLM-judge panel (16 judges, 1,560 scored rows: 104 prompt-matched pools, 5 rolls per model per pool, pooled blind) scored forum-RP continuations at 512-4096 context. Axes 0-2; "paired vs p3" is the mean per-pool difference of the four-axis sum (0-8 scale), each pool comparing 5 rolls against 5 rolls on the same prompt:
| model | consistency | coherence | format | craft | best-pick | paired vs p3 |
|---|---|---|---|---|---|---|
| p3 | 1.61 | 1.32 | 0.88 | 0.94 | 12% | - |
| e1 | 1.62 | 1.37 | 1.02 | 1.01 | 42% | +0.27 (t 3.3) |
| e2 | 1.68 | 1.45 | 0.99 | 1.06 | 45% | +0.42 (t 5.0) |
e2 beats e1 paired (+0.15, t 2.2) despite near-identical losses - the second pass over the selected data moved craft (1.01 -> 1.06) and coherence (1.37 -> 1.45) where the loss curve barely moved; format gave a little back (1.02 -> 0.99). At 4096 - double the training cap, an untrained depth - craft runs 0.97 -> 1.15 and best-picks go 3 -> 11-12 over 26 pools (e1 takes 12, e2 11). In the author's deployment use - one user's observation, not a panel measurement - the e2 merge holds coherent scenes at 8k context, four times its training cap, where the original 5e-6 logbook tunes degraded beyond their training length; consistent with length generalization surviving low-LR doses, on this evidence only.
A later cross-era panel (prompt-matched pools, 16 fresh judges) put e2 at +0.66 paired (t 8.0) over its own base - unusable rate 11.8% -> 1.9%
/gguf/ — F16 and Q8_0 for each of the original three runs, merged into the
base. No GGUFs for the ladder adapters yet.
| file | size |
|---|---|
Ministral-3-14B-v3-lil-{F16,Q8_0}.gguf | 27.0 / 14.4 GB |
Ministral-3-14B-base-lil-{F16,Q8_0}.gguf | 27.0 / 14.4 GB |
Ministral-3-14B-base-clip1-{F16,Q8_0}.gguf | 27.0 / 14.4 GB |
Ministral-3-14B-mixv2-Q8_0.gguf (end of inc 1) | 14.4 GB |
Ministral-3-14B-mixv2p2-Q8_0.gguf (inc 2) | 14.4 GB |
Ministral-3-14B-mixv2p3-Q8_0.gguf (inc 3, era final) | 14.4 GB |
Ministral-3-14B-length-Q8_0.gguf (e1) | 14.4 GB |
Ministral-3-14B-length-e2-Q8_0.gguf (e2, program final) | 14.4 GB |
The per-sample LLM-judge filtering above produced 122k frontier-model verdicts — enough to distill the judging itself. A single-task r32 LoRA on the same base now reproduces the teacher's pass/fail decision at F1 0.80 and, via a probability readout at the quality digit, retrieves 100% of the teacher's top-graded samples within the top 29% of its ranking. It judges the corpus at ~35k windows/hour on one A100, which removed the metered-API ceiling from this project's data pipeline.
Model and full card: Ministral-3-14B-pocket-rp-judge