Contents
- Language-model decoder Linear layers quantized to W4A16 GPTQ (group size 128, symmetric)
- Vision tower + multi-modal projector +
lm_head pass through at original precision (multimodal models)
- Calibrated on training-distribution conversations rendered by
mistral-common
in the same token dialect the model was trained on
Loading (vLLM, Ampere+)
vllm serve BigBlueCeiling/frndo-run-028-warm-restrained-r32-W4A16_GPTQ \
--tokenizer_mode mistral --config_format mistral --load_format mistral \
--tool-call-parser mistral --enable-auto-tool-choice \
--kv-cache-dtype fp8_e5m2
Uses vLLM's gptq_marlin kernel path — works on Ampere (sm_86, RTX A6000/A100),
Ada (sm_89, RTX 4090), Hopper (sm_90), Blackwell (sm_100+). NOT applicable to
Turing or older.
Training focus
Two corpus changes against run-026-depth-max, whose personality is
accepted and deployed. Replies to terse user turns are rewritten warmer across
the four depth corpora, and the tool corpus is taught to offer rather than fire
when the user named an activity instead of the tool's output.
Composition is otherwise r026's exactly. The depth files carry the same
conversations with some assistant turns rewritten, so depth share is unchanged;
only the tool pool grows, which lowers tool oversampling without moving tool
share.
Data composition
Table with columns: Category, Rows, % of run| Category | Rows | % of run |
|---|
tool_calling_training_restrained | 4,000 | 25.0% |
dailydialog_chat_varied | 1,958 | 12.2% |
threats_balanced | 1,714 | 10.7% |
deep_conversations_xl_warm | 1,382 | 8.6% |
|
Tool oversample 1.38x from a 2,896-row pool.
Depth: 4,000 rows from 2,317 unique (1.73x).
Expected behavior
- Terse-arm reply length and warmth should rise. This is the primary
prediction and the terse arm of the conversational bench is the instrument.
- Ambient restraint should appear on
tool_calling_ambient_v1, where r026
is expected to fire on most of the 7 ambient cases.
- Tool positives should hold at 96.2%. A fall here means the model became
timid rather than calibrated, which is the failure this run most risks.
- Warmth overall should rise modestly; only terse-turn replies were touched.
Training-pipeline invariants
Same three invariants as run-005, which every FrndoBrain run inherits:
- mistral-common tokenization only. Every training token sequence is
produced by
mistral-common (Tekken), never by an HF chat template, so it
is byte-identical to what vLLM emits in Mistral tokenizer mode.
[AVAILABLE_TOOLS] on the first user turn, via
UserMessagePosition.first, matched on both the training and serving sides.- Tool-array mixing. Each conversation's
tools array is a random
6-12 tool sample spanning >=2 categories, always including the tools that
conversation actually calls, so the model learns selection under noise
instead of a spurious tools-to-answer correspondence.
Tool-array mixing and tool_call-ID uniqueness are regenerated after the
restraint pass, not inherited, because that pass adds and rewrites
conversations. Oversampling is a full copy of the source with its own shuffle,
not augmentation.
Corpus-quality invariants: 0 slop turns and 0 fabricated-human-life turns.
Rewritten depth turns additionally pass a warmth gate, scored 0-4 by a judge
from a different model family than the rewriter, accepted at 3 with fidelity to
the original's meaning and stance, and labelled friend rather than
therapist or glazing.
Provenance
Quantized by B2CQuantizer. Source: BigBlueCeiling/frndo-run-028-warm-restrained-r32-bf16. Calibration (for calibration-requiring formats) runs on training-distribution conversations rendered via mistral-common with the tools-at-first-user-turn placement invariant enforced.