LoRA adapter for a real-time call-assist model. Given a speaker profile, an objective, a context bundle and a live (speech-recognition) transcript, it emits one natural spoken reply the speaker reads aloud. Adapter only - the base model is not redistributed here.

Base model


Base model	`Qwen/Qwen3.5-9B`
Revision	`c202236235762e1c871ad0ccb60c8ee5ba337b9a`
Architecture	`Qwen3_5ForConditionalGeneration` (dense 9.65B; 32 layers as 8 x (3 x Gated DeltaNet -> 1 x Gated Attention))
Precision	bfloat16

Comparison note. Tokenizer is byte-identical to Qwen/Qwen3.6-27B (verified: 248,077 vocab entries map to identical ids; all 7,764 corpus examples render and tokenize identically), so eval loss is directly comparable across the two models. Pre-training baseline eval loss: 9B 2.935 vs 27B 2.590. Loss parity alone does NOT establish behavioral parity; see Limitations.

Method


Method	LoRA with DoRA enabled (`use_dora=True`)
Rank / alpha	32 / 32
LoRA dropout	0.05
Target modules	12 linear modules, language model only
Trainable params	87,983,616 (0.911% of total)
Trainer	Axolotl 0.18.0 (transformers 5.14.1, peft 0.19.1)
Hardware	1x NVIDIA H100 PCIe 80GB

Target module suffixes: q_proj, k_proj, v_proj, o_proj (Gated Attention); in_proj_qkv, in_proj_z, in_proj_b, in_proj_a, out_proj (Gated DeltaNet); gate_proj, up_proj, down_proj (MLP).

The checkpoint's vision tower is deliberately not adapted (linear_fc1, linear_fc2, qkv, proj). The corpus is 100% text, so those weights would receive no gradient signal; adapting them would add DoRA magnitude vectors that can never be meaningfully updated.

Training data


Train examples	7,219
Eval examples	545 (held out, stratified on `question_type` and `speech_profile`)
Max sequence length	1024 tokens
Effective batch size	16
Epochs run	1.55
Optimizer / LR	adamw_torch, 2e-4 cosine, warmup ratio 0.03

Results

Metric	Value
Final train loss	1.4186
Final eval loss	1.7769
Best eval loss	1.7707 (step 400)
Optimizer steps	700

step	epoch	train loss	eval loss
0	0.00	-	2.9348
100	0.22	1.9009	1.9198
200	0.44	1.8294	1.8446
300	0.66	1.7207	1.8073
400	0.88	1.8188	1.7707
500	1.11	1.3349	1.7977
600	1.33	1.4067	1.7818
700	1.55	1.4186	1.7769

Prompt template

Trained in non-thinking mode. Qwen3.6 is a thinking model whose chat template appends <|im_start|>assistant\n<think>\n on a generation prompt by default; targets here are direct spoken replies with no reasoning trace, so training used an empty think block. The training sequence is byte-identical to the inference prompt produced with enable_thinking=False. Inference must assemble context the same way or quality will degrade silently.

Loss is computed on the response tokens only; all prompt tokens are masked to -100.

System message

You are Jovika, a real-time call assistant. The speaker below is in a live conversation. Produce one natural spoken reply that the speaker can read aloud immediately.

Output only the reply itself. No preamble, no alternatives, no formatting, no bullet points, no em-dashes. Do not open with filler such as OK, sure, yeah, absolutely, certainly, or great question. Lead with the most important point. Write one idea per sentence in short, common words that are easy to say out loud.

SPEAKER PROFILE
Role: {role}
Seniority: {seniority}
Industry: {industry}
Background: {background}

OBJECTIVE
{objective}

DELIVERY
Speech profile: {speech_profile}
Verbosity preference: {verbosity_preference}

If persona.pronouns is present, a Pronouns: {pronouns} line is inserted immediately after the Background: line. Fields id, question_type, domain, messiness and has_uncommitted_tail are deliberately excluded - they are dataset-generation metadata unavailable at inference time, and conditioning on them would create train/serve skew. An uncommitted transcript tail is carried inline in the transcript itself as the literal marker [UNCOMMITTED LIVE TAIL].

User message

CONTEXT DOCUMENTS
{context_bundle}

LIVE TRANSCRIPT
{transcript}

Assistant target

The response string, stripped. Rendered through the base model's own chat template, the full training sequence is:

<|im_start|>system
{SYSTEM}<|im_end|>
<|im_start|>user
{USER}<|im_end|>
<|im_start|>assistant
<think>

</think>

{RESPONSE}<|im_end|>

Reproducing at inference

python

from transformers import AutoTokenizer, AutoModelForImageTextToText
from peft import PeftModel

BASE = "Qwen/Qwen3.5-9B"
tok = AutoTokenizer.from_pretrained(BASE, revision="c202236235762e1c871ad0ccb60c8ee5ba337b9a")
model = AutoModelForImageTextToText.from_pretrained(BASE, revision="c202236235762e1c871ad0ccb60c8ee5ba337b9a", dtype="bfloat16", device_map="auto")
model = PeftModel.from_pretrained(model, "shwnbrns/jovika-qwen3.5-9b-lora-dora")

prompt = tok.apply_chat_template(
    [{"role": "system", "content": SYSTEM}, {"role": "user", "content": USER}],
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=False,   # required: matches training
)

Known data characteristics

Accepted tradeoffs in the corpus, not defects to correct: troubleshooting responses average 5.5 sentences against an intended 8-12 band, and roughly 47% of responses exceed 120 words. The model was trained on the data as given.

Limitations

Not evaluated for output quality against a frontier baseline; only train/eval loss are reported here. The adapter inherits the base model's licence (Apache-2.0).