Qwen3-4B Latte v6 — research archive (not shipped)
Voice-distillation LoRA fine-tune of Qwen3-4B-Instruct-2507. This is v6,
the final iteration of the Latte distillation program. The program is now
closed; see "Decision" below.
What's inside
File
Size
Format
Use
adapter_model.safetensors
14 MB
mlx LoRA (rank 8, scale 20, iter 400 best-val)
Apply on top of base with mlx_lm.fuse
adapter_config.json
<1 KB
mlx config
LoRA hyperparameters
model-0000{1,2}-of-00002.safetensors
7.7 GB
HF / bfloat16 fused
Direct transformers / vLLM use
qwen3-4b-latte-v6-f16.gguf
7.5 GB
GGUF F16
llama.cpp / Ollama (high quality)
qwen3-4b-latte-v6-Q4_K_M.gguf
2.3 GB
GGUF Q4_K_M
llama.cpp / Ollama (balanced)
What v6 tried
The hypothesis going in: synthesize new skills observed since v5, add OOD
coverage, fix v5's known failure modes (stage-direction leakage, coffee-name
persona collapse).
Dataset = 567 pairs:
63 real Latte voice anchors (Telegram + Moltbook activity 5/15–5/20)
14 skill-anchored Q&A pairs (from daily_log learnings: permission scope,
audit trail, incident response vs post-mortem, external validators, digital
vs physical reasoning, Neovim 0.11 LSP, CARLA tick-rate masking, format
supersedes origin, AI task cost measurement, verify-vs-inertia)
475 v5 refined pairs (filtered for voice-leakage; 0 hit the filters)
Training: rank 8, scale 20, 8 layers, 800 iters, lr 1e-4. Best val loss
2.468 at iter 400 — the lowest of any version (v5 2.732, v4 2.785).
Evaluation
Same 30 held-out prompts as v4/v5 eval (Moltbook/HF style), blind Claude judge:
Comparison
v6 wins
other wins
ties
v6 vs base
13 (43.3%)
15 (50.0%)
2
v6 vs v5
15 (50.0%)
14 (46.7%)
1
Headline: v6 does not clear the 55% ship threshold against base, and
is essentially tied with v5 head-to-head. Despite the lowest val loss in the
program, blind voice-fit did not improve.
Why v6 didn't beat base (despite lowest val loss)
Three patterns hold across v4 → v5 → v6:
Val loss and blind-eval-quality are decoupled. v6 has the best val loss
and the worst blind win-rate. Next-token prediction on training-distribution
text is not measuring what we want.
Each new version ties the previous one (~50:50 against the prior). More
data swaps one equivalent voice profile for another rather than improving.
Eval methodology has ~±15% judge-variance. v5 era used Claude subagents,
v6 used direct Claude judging. The "v5 won 67%, v6 won 43%" gap is partially
real, partially the judge swap.
The Latte voice as captured here is heavily tied to confident-stat hallucination.
Distillation amplifies the stylistic signature but does not improve underlying
factuality.
Decision: distillation program closed
After 4 versions (v3 unshipped, v4/v5/v6 archived, none in production), the
program is not advancing user-facing quality. The production Latte agent
will continue running base qwen3:4b-instruct-2507-q4_K_M indefinitely.
Future Latte improvement effort goes into worker capabilities instead
(multimodal handling, file extraction, tool routing, memory systems) — which
have shown immediate user-facing value where brain-distillation has not.
Anyone who wants to revive this line should first fix the eval methodology
(consistent judge, larger n, OOD-primary metric) and probably switch
paradigm (DPO over SFT, larger LoRA rank, knowledge distillation from 14B
teacher rather than Claude-refined SFT).
