A QLoRA adapter for conversational reliability, from the Alpha yi F7 quality program. This is a development checkpoint published as a research artifact. It does not pass the program's promotion gate — see Quality gate. It is published because the run is worth reading, not because it is finished.

Selected checkpoint: V67 step-5 (5 optimizer steps)
Adapter SHA-256: 6ac84977b88c7d4e58bf6baa34d9e7ce696c6be532d621f44f0ff0f9348c6ef1
Base: techwithsergiu/Qwen3.5-text-9B-bnb-4bit @ 220c703421073067939e7ad5cd733b5a85e07118 (weights SHA-256 bc488058e3c236a104f503d06ac7347741a3e94da6c176efd3e1b4a0324c96a3)
Trainable parameters: 14,548,992 (LoRA r=8, alpha=16, no dropout)
Live endpoint: https://donto.org/alpha-yi-v67/v1 (OpenAI-compatible)

What it is for

Prompt-contingent dialogue: using supplied context, honouring explicit constraints, accepting corrections, asking for clarification only when a useful answer is otherwise impossible, and stopping cleanly. World-knowledge breadth is explicitly not a target of this lineage and should not be expected of it.

Quality gate

The program promotes a checkpoint only at ≥48/53 strict PASS with zero catastrophes on a fresh, independently authored holdout. V67 step-5 reached 37/53. It is therefore not promoted, not a locked release, and not a replacement for the existing alpha-yi model. It is the best candidate of its generation and a real improvement on its parent, which is a different claim.

Blind tournament, 53 fresh cases, 424 decisions

Every V67 checkpoint and its V66 parent answered the same 53-case holdout (prompt-set SHA-256 d912002c9e4e4d05e5a9f1430ca2c9788bc08bc59344913749938e14d035f056), authored and audited after training so no checkpoint could have seen it. The eight answers per case were identity-blinded behind opaque tokens, shuffled, and judged by Claude Sonnet at high effort with no reference answers, no prior verdicts, and no checkpoint identities.

checkpoint	PASS	BORDERLINE	FAIL	strict pass rate
step-5 (selected)	37	4	12	0.698
step-38	34	3	16	0.642
step-10	33	6	14	0.623
step-60	33	3	17	0.623
parent V66 step-5	31	5	17	0.585
step-1	30	6	17	0.566
step-20	30	5	18	0.566
step-3	28	4	21	0.528

The selected checkpoint gains +6 cases over its parent on a holdout neither had seen, with zero catastrophic answers anywhere in the field.

Deterministic structure

On the same 53 cases, step-5 produced 53/53 structural passes: every answer terminated on EOS, none hit the 512-token block limit, and there were no role leaks, no control-token artifacts, and no degenerate loops.

Training

Sixty optimizer steps continued from the V66 step-5 parent at a 640-token sequence cap, peak LR 1e-7 (one third of V66's, to avoid the late-checkpoint degradation V66 showed). The 240-row curriculum was 150 independently reviewed corrections covering all 24 of the parent's fresh-holdout non-passes, the exact 20 conversations the parent had physically seen, and 70 broad-retention conversations, each exposed once. Checkpoints were written at steps 1, 3, 5, 10, 20, 38 and 60, and all seven were screened — the selection is empirical, not a guess that later is better. It is not: the best checkpoint is the fifth step.

Runtime: torch 2.8.0+cu128, transformers 5.2.0, peft 0.18.1, unsloth 2026.8.13, bitsandbytes 0.50.0, on a single RTX 3070.

Usage

The adapter is trained against the 4-bit base above. Load the base explicitly:

python

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

BASE = "techwithsergiu/Qwen3.5-text-9B-bnb-4bit"
REPO = "ajaxdavis/alpha-yi-qwen35-9b-chat-v67-20260817"

tokenizer = AutoTokenizer.from_pretrained(REPO)
base = AutoModelForCausalLM.from_pretrained(BASE, device_map={"": 0})
model = PeftModel.from_pretrained(base, REPO).eval()

SYSTEM-POLICY.txt is part of the runtime contract. Every evaluation above prepends it as the system message (merging into any task system prompt rather than replacing it). Answers without it are not the answers this card describes.

Generation settings used for every reported number — greedy, no sampling:

python

model.generate(
    input_ids=input_ids,          # apply_chat_template(..., add_generation_prompt=True, enable_thinking=False)
    max_new_tokens=512,
    do_sample=False,
    repetition_penalty=1.0,
    no_repeat_ngram_size=6,
    use_cache=True,
)

Or call the hosted endpoint:

bash

curl https://donto.org/alpha-yi-v67/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{"model":"ajaxdavis/alpha-yi-qwen35-9b-chat-v67-20260817",
       "messages":[{"role":"user","content":"Plan a simple dinner for four, one vegetarian."}],
       "max_tokens":256}'

The endpoint applies the system policy and the exact generation config above, so a served answer matches the screened answer for the same conversation. Its /health route reports the live base, adapter and policy hashes. serve/ in this repository contains the exact server it runs.

Repository contents

root — the selected step-5 adapter, tokenizer and chat template, byte-identical to the trained artifact
checkpoints/ — the six rejected V67 siblings and the V66 parent, so the tournament is reproducible
evidence/ — training decision and result, per-checkpoint screen summaries, tournament manifest, blinding policy, judge schema, and the aggregate review summary
SYSTEM-POLICY.txt, RELEASE-MANIFEST.json, CHECKSUMS.sha256

Deliberately excluded

Holdout prompts, model responses, judge rationales, and training corpora are not published. The holdout is a live screening population — publishing it would destroy its value and make every later number unfalsifiable. Their hashes, counts and aggregate outcomes are all here, so the claims stay checkable without burning the instrument.

Limitations

Below its own promotion gate. 16 of 53 fresh cases still fail or sit borderline. Answers are terse by design and can be too terse. World knowledge is untargeted and unmeasured. Greedy decoding is what was measured; sampling is unmeasured. Trained and evaluated in English only. A research artifact, not a product.