REN is a 9B model you can interrupt mid-answer.

Evaluated s07 against the base model on heldout_v1 (43 min on 2× T4). At or above base on every headline metric except HumanEval+ pass@1 0.779 vs 0.804 (Δ −0.025, 95% CI [−0.067, +0.012], 8 lost / 4 gained). Gains whose interval excludes zero: format task answer correct +0.337, decide Brier score −0.216, decide accuracy +0.187, format compliance, all kinds +0.106.

Trained s07: 302 steps on 6,875 rows (3.64M tokens) in 4.8 h, with no skipped or NaN steps. Mean training loss 0.73; smoothed final loss 0.69.

Round 7 (s07) started on Kaggle at 21:45 WAT. Data plan: 6,875 rows, 3.64M tokens, 302 steps from ren-data-s07; initialised from ren-s06 @ 6f42694. Token shares: agent 27.7%, coding 24.2%, decide 21.6%, interruption 20.7%, format 5.8%. Its smoke test completed 40 steps (training loss 0.62).

Evaluated s06 against the base model on heldout_v1 (41 min on 2× T4). At or above base on every headline metric. Gains whose interval excludes zero: format task answer correct +0.334, decide Brier score −0.206, decide accuracy +0.177, format compliance, all kinds +0.106. HumanEval+ with the adapter at 0.8× strength: 0.804 vs base 0.804 and full adapter 0.822 (n = 163; 6 lost / 6 gained vs base).

Trained s06: 187 steps on 7,719 rows (2.40M tokens) in 3.4 h, with no skipped or NaN steps. Mean training loss 0.12; smoothed final loss 0.11.

Round 6 (s06) started on Kaggle at 02:09 WAT. Data plan: 7,719 rows, 2.40M tokens, 187 steps from ren-data-s06; initialised from ren-s06-merges @ f9f04ad. Token shares: agent 27.6%, coding 26.7%, interruption 22.4%, decide 16.8%, format 6.5%. Its smoke test completed 40 steps (training loss 0.19).

Evaluated s05 against the base model on heldout_v1 (73 min on 2× T4). At or above base on every headline metric except HumanEval+ pass@1 0.755 vs 0.804 (Δ −0.049, 95% CI [−0.104, +0.006], 15 lost / 7 gained). Gains whose interval excludes zero: format task answer correct +0.283, decide Brier score −0.167, decide accuracy +0.153, terminal task success +0.143, format compliance, all kinds +0.106. HumanEval+ with the adapter at half strength: 0.767 vs base 0.804 and full adapter 0.755 (n = 163; 11 lost / 5 gained vs base).

Trained s05: 740 steps on 20,441 rows (9.51M tokens) in 1.8 h, with no skipped or NaN steps. Mean training loss 0.03; smoothed final loss 0.21.

Round 5 (s05) started on Kaggle at 09:41 WAT. Data plan: 20,441 rows, 9.51M tokens, 740 steps from ren-data-s05; initialised from the base model. Token shares: agent 40.9%, coding 33.7%, interruption 11.7%, decide 10.3%, format 3.4%. Its smoke test completed 40 steps (training loss 0.39).

Smoke test for release stopped before completing, at step 0 of 40 (last progress event: data plan). The instance log reports CUDA out of memory (exit code 5); the full run was not started and the instance powered off.

Evaluated s04 against the base model on heldout_v1 (77 min on 2× T4). At or above base on every headline metric except HumanEval+ pass@1 0.773 vs 0.804 (Δ −0.031, 95% CI [−0.080, +0.018], 11 lost / 6 gained). Gains whose interval excludes zero: format task answer correct +0.323, decide Brier score −0.188, decide accuracy +0.160, format compliance, all kinds +0.103. HumanEval+ with the adapter at half strength: 0.779 vs base 0.804 and full adapter 0.773 (n = 163; 10 lost / 6 gained vs base).

Trained s04: 122 steps on 4,171 rows (1.61M tokens) in 3.2 h, with no skipped or NaN steps. Mean training loss 0.64; smoothed final loss 0.53.

Round 4 (s04) started on Kaggle at 20:29 WAT. Data plan: 4,171 rows, 1.61M tokens, 122 steps from ren-data-v3 (24,172 records); initialised from ren-s03 @ c142b45; LR 2e−5. Token shares: coding 53.9%, agent 16.5%, interruption 15.3%, format 8.2%, decide 6.1%. Phase 0: the base model answered 864 new problems in 48 min; 752 passed the unit tests and became training rows. Its smoke test completed 40 steps (training loss 0.20).

Planned round 4 from the s03 HumanEval+ samples. 14 of the 15 lost problems were already lost by s02, so the drift sits in the inherited s02 adapter; all 15 are wrong implementations (13 failed assertions, 2 timeouts), none are format problems. Base-model self-distillation is the only change with a measured effect (s02 to s03: +4.9 points, 8 of s02's 23 losses recovered), so round 4 trains more of it: 952 reused s03 base answers plus a fresh pool, with coding at about 55% of tokens and the rest replay. Dataset ren-data-v3 (24,172 records); LR 2e−5; continues from the s03 adapter (c142b45). The evaluation adds a HumanEval+ re-run with the adapter at half strength.

Evaluated s03 against the base model on heldout_v1 (83 min on 2× T4). At or above base on every headline metric except HumanEval+ pass@1 0.755 vs 0.804 (Δ −0.049, 95% CI [−0.104, +0.006], 15 lost / 7 gained). Gains whose interval excludes zero: format task answer correct +0.320, decide Brier score −0.181, decide accuracy +0.147, format compliance, all kinds +0.103.

Trained s03: 187 steps on 6,691 rows (2.41M tokens) in 4.6 h, with no skipped or NaN steps. Mean training loss 1.17; smoothed final loss 1.13.

Round 3 (s03) started on Kaggle at 12:13 WAT. Data plan: 6,691 rows, 2.41M tokens, 187 steps from ren-data-v2 (22,065 records); initialised from ren-s02 @ 51de9ea; LR 3e−5. Token shares: coding 29.3%, agent 23.6%, interruption 23.5%, format 13.9%, decide 9.6%. Phase 0: the base model answered 1,280 new problems in 68 min; 1,035 passed the unit tests and became training rows. Its smoke test completed 40 steps (training loss 0.35).

Planned round 3 with ren-data-v2 (22,065 records). Code targets are now the base model's own answers that pass the unit tests (on-policy self-distillation), with no teacher model. Generation stops at <|im_end|> in training and in evaluation, so s03 and later evaluations are not directly comparable with s01 and s02. Added 900 interruption records in which finished work is revised and then resumed, 581 first-action rows and 447 stop records. The run continues from the s02 adapter at LR 3e−5.

Diagnosed the s02 regressions. None of the 23 lost HumanEval+ problems were format failures: 18 failed an assertion, 3 timed out and 2 raised other errors. The likely cause is code targets written by a different model (Qwen2.5-Coder-32B, via OpenCodeInstruct). Separately, Qwen3.5-9B ships without a generation config, so generation did not stop at the end of the turn: 49 of 50 adapter agent answers ran on into invented turns.

Evaluated s02 against the base model on heldout_v1 (92 min on 2× T4). Below base on Agent-turn format compliance 0.880 vs 0.960 (Δ −0.080, 95% CI [−0.180, +0.020], 6 lost / 2 gained); HumanEval+ pass@1 0.706 vs 0.804 (Δ −0.098, 95% CI [−0.166, −0.037], 23 lost / 7 gained); Interrupt handled 0.762 vs 0.810 (Δ −0.048, 95% CI [−0.214, +0.119], 8 lost / 6 gained); Interruption format 0.881 vs 0.929 (Δ −0.048, 95% CI [−0.167, +0.071], 5 lost / 3 gained). Gains whose interval excludes zero: format task answer correct +0.289, decide Brier score −0.138, decide accuracy +0.107, format compliance, all kinds +0.091.

Trained s02: 377 steps on 12,342 rows (4.90M tokens) in 7.0 h, with no skipped or NaN steps. Mean training loss 1.41; smoothed final loss 1.15.

Round 2 (s02) started on Kaggle at 23:12 WAT. Data plan: 12,342 rows, 4.90M tokens, 377 steps from ren-data-v1 (18,039 records); initialised from the base model; LR 5e−5. Token shares: coding 33.3%, agent 28.0%, interruption 20.7%, format 9.4%, decide 8.7%. Its smoke test completed 40 steps (training loss 0.59).

Built ren-data-v1 (18,039 records) to address the pilot regressions. Removed Llama-generated records. Added 3,480 code-only rows, 1,396 interrupt records rendered with the evaluation's tool prompt, and 300 strict first-action rows. Raised the token budget from 4.08M to 4.90M.

Pilot samples: 23 of 163 code answers opened with prose, and failed stop requests replied “Stopped.” and then made another tool call or invented a user turn.

Evaluated s01 against the base model on heldout_v1 (79 min on 2× T4). Below base on Agent-turn format compliance 0.900 vs 0.960 (Δ −0.060, 95% CI [−0.160, +0.040], 5 lost / 2 gained); HumanEval+ pass@1 0.675 vs 0.804 (Δ −0.129, 95% CI [−0.202, −0.061], 27 lost / 6 gained); Interrupt handled 0.524 vs 0.810 (Δ −0.286, 95% CI [−0.500, −0.071], 19 lost / 7 gained); Interruption format 0.714 vs 0.929 (Δ −0.214, 95% CI [−0.381, −0.048], 12 lost / 3 gained). Gains whose interval excludes zero: format task answer correct +0.303, format compliance, all kinds +0.097, decide Brier score −0.095.

Trained s01: 321 steps on 8,617 rows (4.08M tokens) in 6.0 h, with no skipped or NaN steps. Mean training loss 2.23; smoothed final loss 1.85.

The pilot (s01) started on Kaggle at 22:23 WAT. Data plan: 8,617 rows, 4.08M tokens, 321 steps from ren-data-v0; initialised from the base model; LR 5e−5. Token shares: agent 58.6%, coding 10.8%, interruption 10.4%, decide 10.4%, format 9.8%. Its smoke test completed 40 steps (training loss 0.60).