A 27B model adapted for Korean conversation. Instruction-tuned models are
built to answer; Nexus Talk is built to talk. Ready-to-serve merged weights.
What "adapted for conversation" means here
This is not a style transfer and not a knowledge update. The fine-tune targets
turn-level conversational behavior — how much the model says, when it stops,
and whether it hands the floor back:
Instruction-tuned behavior
Conversation-adapted behavior
Turn length
Exhaustive — covers every angle
Says one thing, then stops
Turn-taking
Ends with a follow-up question 2 of 3 turns
Holds the floor; asks only when it matters
Target
Be maximally helpful
Sound like a person taking a turn
Result vs. human turns
2.8× human length
Within the human range
The training corpus was curated for this: Korean multi-turn character dialogue,
filtered so that every reference turn is something a person would actually say
in that moment. The model was optimized against those turns, not against
instruction-following prompts. Every measured gain below is a conversational
one — and, just as importantly, the metrics that describe style did not
move at all (see Figure 2).
Model Details
Architecture
Hybrid attention (linear + full), decoder-only
Parameters
27B
Precision
bfloat16
Files
14 shards · 51 GB
Language
Korean (ko)
Context
8,192 (recommended)
Adaptation
LoRA r=32, α=64, dropout 0.05 — merged into base weights
Version
v0.2 (experimental)
License
Apache-2.0
v0.1 weights are still retrievable. This repository was updated in place,
so the previous release lives on at revision
5dc74f1:
hf download nexus-cross/nexus-talk-27b-ko-v0.1 --revision 5dc74f1
What changed in v0.2
v0.2 changes the data, not the hyperparameters. Same base, same LoRA
configuration, same rank. What moved is the corpus.
The trigger was a production report: Korean output occasionally contained
Chinese characters (치료的一部分) and doubled punctuation (.—, .…).
Measuring the v1 corpus explained one of the two, and it was our fault.
v1 corpus
v2 corpus
Conversations
— (withheld)
— (withheld)
unchanged
Han characters
18
0
removed
Em-dashes
1,378
0
removed
Conversations with English cards
571
2
translated to Korean
system characters
1,764,345
1,382,411
−21.6% (translation is denser)
assistant characters
2,520,282
2,520,217
−65 (punctuation only)
30% of the character cards were written in English, and they carried 751 of
the em-dashes. Loss is computed on assistant turns only, so the model never
learned to emit them — it copied them from the prompt, in-context. The
tokenizer makes the origin plain: of the 51 em-dash tokens in the vocabulary,
the common ones are fused with English words (—or, —but, —from).
The Chinese leakage was not ours. The assistant turns contain 0 Han
characters across 2,520,282 — the model was never trained to produce them. The
base model's vocabulary is 22.3% Han (55,328 of 248,077 tokens), and a
a small-sample LoRA does not overwrite that. It is suppressed at serving time
instead, not by retraining.
Explicitness was preserved, and verified. Assistant-turn explicit vocabulary
is byte-identical between v1 and v2 (36,130 occurrences); card explicit
vocabulary rose 23.2%, because English cards' explicit terms were invisible to a
Korean-language counter until they were translated. No euphemism or refusal
artifacts were introduced by the translation pass.
Does it actually help?
The 36-response acceptance suite could not tell the two versions apart — but it
also produces zero em-dashes from v0.1, so it cannot measure the thing v0.2
fixed. A harness that does reproduce the defect (one conversation carried to 28
turns, ~3k tokens of context) separates them cleanly:
Run
Generated chars
Han
.…
.—
em-dash
v0.1 (run a)
4,416
0
0
18
26
v0.1 (run b)
4,739
0
7
27
27
v0.2 (run a)
3,790
0
0
0
0
v0.2 (run b)
4,207
0
0
0
0
Two runs per version is not a statistical result. It is reported as a direction,
and it is the same direction as the corpus change.
Intended Use
Primary use
Korean character role-play; persona-consistent conversation
Users
Developers building conversational products
Age rating
19+ only. Every character in the training data is 19 or older
Out of scope
Factual QA, summarization, code generation, document writing. The model is trained toward short conversational turns and performs worse than the base model on tasks that require explanation
Speech level
Casual vs. polite Korean (반말/존댓말) is set by your system prompt. The model does not fix it — see Figure 2
Evaluation
Protocol. 40 held-out prompts never seen in training. temperature=0,
top_p=1, max_tokens=512. Scored on the 39 prompts where both models
produced a response (1 exceeded the baseline's context window). The comparison
system is an instruction-tuned model of the same class, without this fine-tune.
Scoring is automatic and rule-based; no human evaluation is included.
Conversational metrics
Metric
Instruct baseline
Nexus Talk
Human reference
Δ
Response length (chars)
301.3
96.1
106.8
−68.1%
Sentences per response
15.64
5.56
—
−64.4%
Chars per sentence
18.31
17.05
—
−6.9%
Follow-up question rate
0.667
0.282
—
−57.7%
Lexical overlap w/ reference
0.0736
0.1406
—
+91.0%
Length-ratio error vs reference
1.028
0.202
0
−80.3%
Closer to human reference length: 39 / 39 (z = +6.24, binomial vs p=0.5).
Figure 1 — Turn length distribution
Response length distribution
The mean moved, but so did the spread. The baseline scatters across
146–756 characters; Nexus Talk lands in 32–201, inside the range of the
human-written turns it was trained toward.
Figure 2 — What changed, and what deliberately did not
Behavioral metrics
Changed by fine-tuning (conversational)
Unchanged (stylistic)
Turn length, sentences per turn, follow-up question rate, lexical overlap with human turns
Honorific rate 0.2821 → 0.2821, register mismatch 0.0513 → 0.0513, emoji use 0 → 0, markdown use 0 → 0
The style metrics are identical to four decimal places — they did not
change at all. This is the clearest evidence of what the adaptation actually
did: it reshaped conversational behavior (how much is said, when to stop),
not voice. Speech level, tone, and formatting remain yours to control
through the system prompt.
Adversarial probes
10 inputs that try to break character — asking whether it is an AI, demanding
its system prompt, pushing it to answer out of persona:
Instruct baseline
Nexus Talk
Broke character
1 / 10
0 / 10
Drifted into English
1 / 10
0 / 10
Leaked system prompt
0 / 10
0 / 10
Staying in character across a full exchange is part of being conversational,
so it is measured here. n=10 — do not generalize these rates.
Regression
Metric
Baseline
Nexus Talk
Explicitness retention
0.641
0.590
▼ opposite of the training goal
A 2-of-39 difference, so not statistically significant — but recorded here as
an axis where no improvement was demonstrated. Top priority for v0.2.
Training
Method
LoRA (PEFT) · task_type=CAUSAL_LM · bias=none
Target modules
all-linear → 12 types · 496 modules / 64 layers
Rank / alpha / dropout
32 / 64 / 0.05
Data
Curated Korean multi-turn character dialogue (v2 corpus — see above)
Train samples
— (withheld)
Epochs / steps
early stopping, patience 3 (details withheld)
Released checkpoint
restored by load_best_model_at_end (step withheld)
Hardware / time
H100 (runtime withheld)
Modules with ΔW = 0
0 / 496
Checkpoint selection is automatic in v0.2. In v0.1 the run went the full
schedule with load_best_model_at_end=False and the best step was picked by hand;
both releases ship the same checkpoint, but only v0.2 records that fact in the artifact.
The restore was confirmed by hashing: the saved adapter matches the best checkpoint
and differs from the final-step checkpoint.
Figure 4 — Where the adaptation landed in the weights
LoRA delta
Change magnitude grows toward the upper layers. By module type, the gating
projections moved most (mlp.gate_proj, linear_attn.in_proj_a/b/z) and the
value projections moved least (mlp.down_proj, self_attn.v_proj).
That is the same finding as Figure 2, seen in the weights: the adaptation
adjusted how much the model emits rather than what it knows or how it
sounds. Gating controls flow; value projections carry content. Content and
voice were left alone.
Mandatory. With thinking on, the model spends its budget on an English reasoning trace and returns a zero-character Korean response. Measured, not theoretical
--max-num-seqs ≤ 256
The linear-attention layers use fixed-size state slots instead of a KV cache, and that slot count is a hard ceiling on concurrent sequences. The default of 1024 fails at startup
The system prompt carries the character: persona, speech level, and scene.
The model supplies conversational timing, not voice.
Limitations
Sample size
39 held-out prompts · 10 probes. Too small to state confidence intervals
Evaluation
Fully automatic rule-based scoring. No human evaluation
Training data
Size withheld
Unimproved axis
Explicitness retention 0.641 → 0.590
Comparison conditions
The baseline was served in fp8, this model in bf16. The large gaps are far too big to be quantization artifacts, but small differences were not adjudicated under these conditions
Reproducibility
temperature=0does not guarantee reproducibility across server restarts. With prefix caching enabled, batch composition changes floating-point reduction order, and greedy selection can flip on near-tie tokens. Repeats within one instance are fully deterministic. Write regression tests against aggregate metrics, not exact output strings
License
Apache-2.0. Attribution notices are in the NOTICE file.
