arabic-rag-chat-8k-eval

Per-row evaluation artifacts for the 8,192-token Arabic multi-turn RAG models: the test split, every model's raw replies, every judge verdict, and the rendered report for each. Thirteen judged models, all scored on the same 1,651 prompts by the same judge at temperature 0.0, so the comparison below is like-for-like and can be recomputed offline without a GPU or a judge server.

This is the measurement half of oddadmix/100M-8192-Nawah-dsv4; the training corpora are oddadmix/arabic-rag-chat-30K and oddadmix/arabic-rag-chat-grpo-5K.

Layout

path	rows	what it is
data/test.jsonl	632	the test split — 632 conversations, company-disjoint from train, expanding to 1,651 gold-forced turn prompts
answers/<tag>.jsonl	1,651	one model's replies. {"id": "<conv>-<bucket>-<n>#r<round>", "reply": "..."}
judged/<tag>.jsonl	1,651	judge verdicts. {"id", "answerable", "verdict": {"refused", "score", "hallucination"}}
reports/<tag>.txt	—	the rendered report: overall, per length bucket, per round index

id is the join key across all three. score is 0 (wrong or hallucinated), 1 (partially correct), or 2 (correct and fully grounded).

Reproducing a report

No GPU, no judge server — the verdicts are already here:

python eval-rag-support.py report --multiturn \
    --test-file data/test.jsonl --tag chat-grpo-v3

eval-rag-support.py ships in the training/ folder of the model repo. Place answers/<tag>.jsonl and judged/<tag>.jsonl as rag-support-eval/answers-<tag>.jsonl and rag-support-eval/judged-<tag>.jsonl, which is where the script looks.

Gold-forced replay

Each conversation is expanded into one prompt per round, with rounds 0..k−1 replayed from the gold answers rather than the model's own. A round-4 score therefore measures round 4, not the damage done in round 2. This isolates per-turn quality from error cascade — it is deliberately the easier of the two measurements, and a free-running multi-turn eval would score lower.

Results

Ranked by judge score. All figures from reports/, 1,651 turn prompts each.

tag	score	halluc	part-cov	agg-recall	distractor-num	refusal (unans.)	chrF++
chat-grpo-v3	0.86	29.1%	58.5%	53.7%	20.7%	139/149 (93%)	46.93
chat-grpo-v2	0.81	37.1%	58.0%	56.5%	21.1%	141/149 (95%)	47.51
chat-grpo	0.77	39.0%	57.1%	56.6%	19.8%	141/149 (95%)	46.53
chat-fullattn-s1	0.73	37.9%	49.6%	51.4%	24.1%	143/149 (96%)	43.32
chat-s1	0.72	38.1%	50.1%	53.0%	22.4%	141/149 (95%)	43.75
chat-lfm2s-grpo	0.30	54.8%	24.4%	21.9%	32.5%	77/149 (52%)	26.33
chat-2k-lfm2-grpo	0.27	53.8%	23.5%	21.2%	34.1%	78/149 (52%)	24.41
chat-lfm2s	0.27	57.5%	20.1%	19.5%	28.3%	105/149 (70%)	24.38
chat-2k-lfm2	0.25	57.2%	18.9%	16.5%	28.1%	128/149 (86%)	23.65
chat-2k-grpo	0.24	37.1%	17.4%	14.4%	8.2%	50/149 (34%)	15.60
chat-2k	0.21	40.1%	14.7%	12.7%	16.6%	69/149 (46%)	14.57
chat-2k-r	0.21	42.3%	14.8%	12.8%	22.0%	57/149 (38%)	14.51
chat-rag2k	0.03	61.0%	5.3%	5.6%	27.2%	66/149 (44%)	11.91

The same models by length bucket

The overall column above is misleading on its own, and this table is why:

tag	long	mid	short	chrF++ long → short
chat-grpo-v3	0.81	0.93	0.84	47.97 → 45.47
chat-grpo-v2	0.77	0.82	0.84	48.56 → 46.82
chat-grpo	0.74	0.81	0.76	47.96 → 44.60
chat-s1	0.71	0.72	0.73	44.94 → 42.19
chat-fullattn-s1	0.66	0.74	0.79	44.45 → 42.30
chat-lfm2s-grpo	0.03	0.07	0.79	15.23 → 45.17
chat-2k-lfm2-grpo	0.02	0.05	0.73	14.12 → 42.45
chat-lfm2s	0.02	0.05	0.73	13.34 → 41.53
chat-2k-lfm2	0.03	0.05	0.67	13.03 → 39.95
chat-2k-grpo	0.00	0.00	0.72	6.89 → 41.59
chat-2k	0.00	0.00	0.62	7.78 → 38.85
chat-2k-r	0.00	0.00	0.61	7.29 → 38.90
chat-rag2k	0.00	0.00	0.10	8.45 → 22.48

Columns. score — mean judge score 0–2. halluc — share of answered turns the judge flagged as ungrounded. part-cov — fraction of each turn's teacher-supplied key_facts present in the reply (judge-free string test; every key_fact was validated to occur verbatim in its cited passage). agg-recall — over turns needing ≥2 passages, the fraction of gold passages that contributed at least one fact. distractor-num — figures quoted from a non-gold passage. refusal — unanswerable turns correctly declined.

Which model each tag is

tag	model	notes
chat-s1	Nawah-50M-RAG-Chat-8K-S1-FINAL	50M Gemma-3, 8K ctx, SFT stage 1
chat-s2	Nawah-50M-RAG-Chat-8K-FINAL	SFT stage 2 (anneal) — answers only, never judged
chat-fullattn-s1	Nawah-50M-RAG-Chat-8K-FullAttn-S1-FINAL	as chat-s1 with sliding window disabled
chat-grpo	Nawah-50M-RAG-Chat-8K-GRPO-FINAL	chat-s1 + GRPO, v1 rewards
chat-grpo-4400	Nawah-50M-RAG-Chat-8K-GRPO step 4400	mid-run checkpoint — answers only, never judged
chat-grpo-v2	Nawah-50M-RAG-Chat-8K-GRPO-v2-FINAL	v2 rewards
chat-grpo-v3	Nawah-50M-RAG-Chat-8K-GRPO-v3-FINAL	v3 rewards (+gold_coverage)
chat-2k	Nawah-50M-RAG-Chat-2K-FINAL	50M Gemma-3 at 2K ctx — context ablation
chat-2k-r	Nawah-50M-RAG-Chat-2K-R-FINAL	2K, retrieval-reranked variant
chat-2k-grpo	Nawah-50M-RAG-Chat-2K-GRPO-FINAL	chat-2k + GRPO
chat-2k-lfm2	Nawah-100M-RAG-Chat-2K-FINAL	100M LFM2, 2K ctx
chat-2k-lfm2-grpo	Nawah-100M-RAG-Chat-2K-GRPO-FINAL	+ GRPO
chat-lfm2s	Nawah-100M-RAG-Chat-2K-S-FINAL	100M LFM2, staged SFT
chat-lfm2s-grpo	Nawah-100M-RAG-Chat-2K-S-GRPO-FINAL	staged SFT + GRPO
chat-rag2k	Nawah-50M-RAG-Support-FINAL	older single-turn support model, run out of domain

Each attribution was read back out of the generation log or the chain script that produced it, with one exception: no generation log survives for chat-grpo-v2, whose model directory is assigned by the naming convention that chat-grpo and chat-grpo-v3 were both confirmed to follow.

Reading the table honestly

The overall score is not a model ranking — it is a truncation measurement. Every 2K model scores exactly 0.00 on long and mid: for the three Gemma-3 ones, none of the ~510 graded long turns or ~488 graded mid turns scored above 0, with chrF++ collapsing to 4–9. Their prompts are cut to 2,048 tokens, so the passage holding the answer is usually not in the input at all and the model is completing a fragment. The 8K-vs-2K gap in the headline table is therefore almost entirely how much of this test set exceeds 2K — 66% of turns — and almost not at all how good those models are.

On the bucket that fits, the 2K models are competitive. Restricted to short, where the whole conversation is inside a 2,048-token window, chat-lfm2s-grpo scores 0.79 against chat-s1's 0.73 — a 100M LFM2 model at 2K beating the 8K Gemma-3 SFT it is nominally a baseline for, and matching chat-fullattn-s1. Only the GRPO'd 8K models (0.84) stay ahead. Any claim that long context is what makes these models work has to survive that column.

chat-2k-grpo's 8.2% distractor rate is not a win either. It is not refusing — it answers 98% of answerable turns. It scores 8.2% because on long and mid it emits almost nothing gradeable (chrF++ 6.89 and 4.13), and text that quotes no figures cannot quote a wrong one. Read that column against chrF++ and the per-bucket scores, never alone.

chat-rag2k is doubly out of domain: a single-turn 2K support model replaying multi-turn prompts. Its 0.03 is a sanity floor, not a comparison.

GRPO moved coverage, not aggregation. Across three reward revisions, part-coverage went 50.1 → 58.5 and chrF++ 43.75 → 46.93, while aggregation-recall stayed at 53.0 → 53.7 and actually peaked at v1 (56.6). The GRPO gains are in saying more of the right things per passage, not in combining more passages — and multi-passage aggregation is precisely what an 8K window was supposed to buy.

The full-attention ablation is null overall but not per bucket. Removing the sliding window moved the headline score by +0.01 (0.72 → 0.73), which looks like noise. Per bucket it traded: long 0.71 → 0.66 and mid 0.72 → 0.74, short 0.73 → 0.79. Giving every layer the full window made the model worse on exactly the bucket it was supposed to help. That is one run and the buckets are ~550 rows each, so it is suggestive rather than settled — but it is not the "no effect" the overall number implies, and it is the result most worth reproducing before betting further on attention span.

Stage 2 has no row. chat-s2 answers exist but were never judged, so the stage-1-vs-stage-2 question — the two-stage token-mix correction the SFT is built around — is currently unmeasured. Its answers are shipped here so that judge pass costs only judge time.

Judge

gemma-4-31B-it behind an OpenAI-compatible endpoint, temperature 0.0, classifying each reply for refusal, correctness (0/1/2), and grounding. The judge sees the gold answer and the cited passages. Judge-free columns (part-coverage, aggregation-recall, distractor-number, chrF++) are computed by string and n-gram tests and do not depend on it.

All generation ran on a single consumer GPU.

Limitations

The test set is teacher-generated synthetic Arabic customer-support conversations, not production traffic. The corpus forbids arithmetic by construction — the teacher quotes figures and never computes new ones — so nothing here measures numerical reasoning. Judge scores from a 31B model on outputs of 50–100M models are coarse; the judge-free columns are the more reliable signal and are the ones to prefer when the two disagree.

© KAND CA 2026