etpi-phase1 — Structured-CoT SFT seed for Qwen 3.6 27B
etpi-phase1 is Qwen 3.6 27B with one round of supervised fine-tuning
on grammar-constrained, structured-CoT coding-agent trajectories. The
adapter (LoRA rank 128, target_modules=all-linear) has been merged into
the base weights — this is a normal HuggingFace model checkpoint with
no PEFT runtime dependency.
This is phase 1 of a multi-phase training pipeline. Phase 2 is RLVR
(GRPO) against R2E-Gym, starting from this checkpoint.
What's different about this model
The base Qwen 3.6 27B reasons in long free-form chains-of-thought
(hundreds-to-thousands of tokens per turn). This SFT seed teaches the
model a compact 3-slot IR inside <think>...</think>:
<think>
STATE: <one short line: current workspace state>
ACTION: <one short line: what to do next and why>
EXPECT: <one short line: what the observation should look like>
</think>
<tool_call>...</tool_call> # or whatever the harness expects
Typical thinking-token spend per turn drops from ~1000+ → ~50-100
while preserving multi-turn coherence and tool-use ability.
Intended use
Drive a multi-turn coding agent — e.g., R2E-Gym / SWE-bench /
Terminal-Bench style sandboxes — where the model:
Reads a task instruction (GitHub issue, terminal prompt, etc.)
Issues tool calls in a loop (bash, file editor, search)
Observes results and iterates
Submits when verified
The IR structure compresses the thinking phase; tool calls and answers
are left unconstrained.
LoRA r=128, α=256, dropout 0.05, target_modules=all-linear, merged at end of training
Optimizer
AdamW (8-bit), lr 2e-4, cosine schedule, warmup 0.03
Epochs
2
Batch
per-device 1, grad accum 8 (effective batch 8)
Sequence length
8192
Loss masking
assistant tokens only (observations and user messages masked)
Other
gradient checkpointing on, bf16, Liger fused CE loss
Hardware
1× H100 80GB (~40 min)
Final metrics:
train_loss (final): 0.096
mean_token_accuracy: 95.87%
Data provenance
Trajectories were generated by running grammar-constrained
Qwen 3.6 27B as a multi-turn agent against R2E-Gym-Lite (real GitHub
issue tasks). The grammar enforced the IR structure inside <think>
while leaving tool calls free. Each task was rolled out 4 times (K=4
best-of-N). Only reward=1.0 trajectories (passed the R2E unit-test
verifier) were kept; among those, the shortest trajectory per task
was selected as the SFT target — encoding a brevity-given-correctness
preference directly into the data.
See andthattoo/etpi-sft
for the full dataset including system prompt, task instructions, and
per-turn messages.
Limitations
Small training set (318 examples). SFT-seed scale, not full SFT.
Expected to generalize the IR format rather than acquire new
capabilities. The phase-2 RL run is where capability climbs.
Trained on one task distribution (R2E-Gym, Python SWE-style
issues). Performance on other languages or task types is untested.
R2E paper authors use the Terminus 2 scaffold with an 80k-token-
per-turn budget to report 77.2% on SWE-bench Verified for the base
Qwen 3.6 27B. This model is intentionally trained under a tighter
scaffold (bash-only loop, IR-constrained thinking) with a different
efficiency objective — direct numerical comparison is not apples to
apples.
Phase 1 only. No RL has been applied. The expected pass@1 lift on
R2E-Gym from SFT alone is modest; the real lift comes from phase 2.
Recommended inference
The model has internalized the IR format. You can drive it with or
without grammar enforcement at inference time:
