We run reinforcement learning (RL) with GRPO across four environments, starting from a supervised fine-tuned 30B-A3B mixture-of-experts model covering English and eight Indic languages. We compare the model before and after RL across 34 benchmarks in one evaluation campaign. MATH improves by 36.07 points, and most Indic benchmarks also improve. Yet XQuAD-Indic F1 falls 19.32 points as English answers rise from 8.7% to 91.7%; IndicQA falls 8.64, while English XQuAD gains 11.71 F1. A second RL run from the same supervised initialization adds an Indic environment together with language and repetition rewards. Compared with the released model, it improves both passage-answering and instruction-following benchmarks, along with preserving MATH performance. But it loses ground on reasoning and coding benchmarks. Several training choices changed together, so we cannot isolate the effect of any one of the reward terms. Script compliance remains unresolved. Adding a language reward did not ensure use of the requested script. Passage answering and script compliance therefore need direct evaluation, even when other benchmark scores improve.
Figures & tables
SFT-Init Shared supervised initialization
↙ separate RL runs ↘
RL-Base Released checkpoint RL-Indic Separate research run Campaign A: SFT-Init vs. RL-Base Campaign B: RL-Base vs. RL-Indic
Figure 1: Both RL runs start from SFT-Init ; RL-Indic is not a continuation of RL-Base . The campaign labels denote evaluation comparisons, not training stages.
Campaign A
Campaign B
Benchmark
SFT
RL-Base
Δ
RL-Base
RL-Indic
Δ
English controls
MATH
48.69
84.76
+36.07
84.76
86.76
+2.00
IFEval
66.36
82.81
+16.45
82.81
86.88
+4.07
XQuAD-English (EM)
46.97
63.70
+16.73
63.70
61.85
−1.85
Indic passage answering
Table 1: Selected results from two separate comparisons. Campaign A measures the change from SFT-Init to RL-Base . Campaign B measures the change from RL-Base to RL-Indic . Each Δ is measured within its own campaign, in points. No Δ reported in this paper is computed across campaigns. Bold highlights the main gains and declines. EM denotes exact match, scored under one reconciled extraction rule across all three checkpoints. Because that rule differs from the one used in earlier records, the text reports passage answering in F1, which is unaffected; Appendix D carries both. All instruction-following scores use prompt-strict accuracy.
MATH
+36.07
MATH-500
+35.20
AIME 2025
+1.46
AIME 2026
+2.50
HMMT February
+1.25
Figure 2: Selected benchmark changes in points. Bars are proportional within each gains or declines panel; scales differ between panels, and metrics are benchmark-native. Deltas are computed within each campaign only.
Appendix figures & tables9 assets
Supplementary material from the paper’s appendix.
Appendix
Setting
Value
Export point
partway through the final pass, verified from export metadata
Scheduled sequence positions
80.53B ( 2,400×256×131,072 ), including padding
Approximate loss-bearing positions
∼ 15B (18.2%, assistant-only mask)
Trainable-rate measurement
95,515 / 524,288 positions over four packed rows
Intra-pack packing efficiency
99.98%
Prepared packed capacity
92.12B positions across 702,824 sequences
Appendix
Table 2: Training record for SFT-Init , the supervised checkpoint used to initialize both RL runs.
Benchmark
Metric
Final SFT
RL ckpt
Δ (RL − SFT)
ckpt
step 40
Math & Reasoning
MATH
exact match
48.69
84.76
+36.07
MATH-500
exact match
48.80
84.00
+35.20
GSM8K
exact match
85.75
94.31
+8.56
MGSM-English
exact match
91.60
98.40
+6.80
Appendix
Table 3: Campaign A: SFT-Init and the released RL-Base checkpoint, evaluated with the same harness, model mode, and decoding configuration. HumanEval+ is omitted. Each IndicIFEval-ground row combines identical prompt-level and instruction-level scores. Changes are RL minus SFT. Accuracy metrics are percentages. The chrF++, chrF, and BLEU values are sacreBLEU scores on a 0–100 scale, not percentages. For rows marked ↓ , lower scores are better, so a negative change indicates improvement. Bold marks the better score in each row.
Table 4: Released RL-Base configuration. Each environment returns its own binary or task-specific reward. No additional reward shaping is applied.
Setting
Value
Pool
Rows
Share
Added environment
indic_if
Mathematics
6,192
40.5%
Reward
Rtask+0.3Rlang−0.0Rlen−0.3Rrep
Indic instruction following
3,179
20.8%
Malformed / abandoned
−0.25
Structured outputs
1,932
12.6%
Advantage std. floor
0.05
Workplace assistant
1,804
11.8%
Steps / rows / validation
40 / 15,300 / 256
English guardrail
1,509
9.9%
Recorded rollouts
251,906
MCQA
684
4.5%
Appendix
Table 5: RL-Indic configuration and input mixture. The rows are filtered by difficulty and then capped within each environment. Shares describe the resulting pool, not preset mixture weights. Rounded shares sum to 100.1%, while the row counts sum to exactly 15,300.
Environment
we
dˉe
contribution wedˉe
Structured outputs
43.4%
+0.2588
+0.1123
Workplace assistant
19.3%
+0.2483
+0.0480
MCQA
7.5%
+0.1030
+0.0077
Mathematics with judge
29.8%
+0.0006
+0.0002
Aggregate dˉ
+0.1681 (SE 0.0095)
95% bootstrap CI
[+0.1489,+0.1871]
Appendix
Table 6: Paired reward differences in the retained rollout sample. Each environment contributes its mean difference weighted by its share of paired prompt groups. Here we is that share within these RL-Base rollouts, not the environment’s share of training rows in the RL-Indic input mixture. Uncertainty is reported as a standard error (SE) and a confidence interval (CI).
Benchmark
SFT
RL-Base
ΔBase
RL-Indic
ΔIndic
Math & Reasoning
MATH
48.69
84.76
+36.07
86.76
+38.07
MATH-500
48.80
84.00
+35.20
87.60
+38.80
GSM8K
85.75
94.31
+8.56
91.58
+5.83
MGSM-English
91.60
98.40
+6.80
97.60
+6.00
MGSM-Indic
79.60
84.60
+5.00
86.40
+6.80
Appendix
Table 7: Three-way matched comparison over the 34 benchmarks scored in both campaigns. The RL-Base column is identical, value for value, to the corresponding column of Table 3 ; that agreement across all 34 is our evidence that the two campaigns share an evaluation state, and is corroboration rather than a provenance record. Metrics are benchmark-native and are the single reported metric per benchmark where several exist; the complete 55-metric SFT-Init / RL-Base pair is Table 3 .
Benchmark
RL-Base
RL-Indic
Δ
MultiChallenge
0.4925
0.5000
+0.75
TauBench airline
0.3733
0.4400
+6.67
TauBench retail
0.1520
0.2807
+12.87
TauBench telecom
0.0936
0.3626
+26.90
SWE-bench Verified
0.2280
0.1220
−10.60
Appendix
Table 8: Tool-use and multi-turn benchmarks for RL-Base and RL-Indic . Scores are on a 0–1 scale. Changes are percentage points, with positive values favoring RL-Indic . Bold marks the higher score in each row.
Category
SFT-Init
RL-Indic
Headline, renormalized over 17 categories
53.81
51.16
Parallel / parallel-multiple
69.5 / 80.0
0.5 / 0.5
Live parallel / live parallel-multiple
75.0 / 75.0
0.0 / 0.0
Simple Python / multiple
86.8 / 81.0
94.5 / 93.0
Multi-turn base
34.0
58.0
Appendix
Table 9: BFCL v4 scores for SFT-Init and RL-Indic . The overall score falls by 2.65 points, with the largest losses in categories requiring parallel tool calls.
Campaign
Ground strict
Translated strict
Ground in-script
Initial released-run matched campaign
28.66
47.45
5.4%
Dedicated IndicIFEval campaign
63.88
—
54.2%
Paired quantization campaign, BF16 baseline
33.05
49.87
—
Appendix
Table 10: RL-Base scores across three IndicIFEval evaluation campaigns. Dashes mark measurements not retained in the available record. The dedicated campaign records 65.62 under the loose criterion. The paired quantization campaign supplies the campaign-A IndicIFEval-ground scores in Table 1 ; the other two rows enter no reported delta. The differences between campaigns remain unresolved, so comparisons require a matched evaluation setup.