SFT-trained PEFT adapter for incident triage on top of Qwen/Qwen2.5-3B-Instruct. Same training objective as the 1.5B adapter: improve structured triage quality across diagnosis, policy, blast radius, and PR-quality heads using oracle trajectories from the OpenEnv incident-triage environment.

Note on the repo name: This model ID uses grpo-train for historical alignment with the hackathon workflow; the published weights are the SFT stage unless you separately run GRPO refinement (see project README).

Model Details

Developed by: OpenEnv Hackathon project team
Model type: Causal LM adapter (LoRA / PEFT)
Base model: Qwen/Qwen2.5-3B-Instruct
Language: English
License: Apache-2.0 (inherits base model compatibility requirements)
Frameworks: Transformers + PEFT
Rough footprint (bf16 inference): ~6.5 GB VRAM with base + adapter; ~8.4 M trainable params, ~33 MB adapter on disk (see project README)

What this model is for

infer likely root cause from noisy multi-signal incident context
recommend policy-level response actions
estimate blast radius and urgency
generate PR/remediation guidance in a concise format

Intended and out-of-scope use

Intended use

internal incident simulation and benchmarking
triage copilots for engineering on-call workflows
structured response drafting where a human reviewer remains in-the-loop

Out-of-scope

fully autonomous production incident response without human approval
legal/compliance sign-off
medical, financial, or safety-critical decision automation

How to run

python

from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel
import torch

base_id = "Qwen/Qwen2.5-3B-Instruct"
adapter_id = "AbhishekMallick/incident-triage-grpo-train-Qwen3B"

tokenizer = AutoTokenizer.from_pretrained(base_id, trust_remote_code=True)
base_model = AutoModelForCausalLM.from_pretrained(
    base_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True,
)
model = PeftModel.from_pretrained(base_model, adapter_id)

prompt = "Incident: API latency spike after region failover. Provide diagnosis, policy, blast radius, and PR plan."
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
with torch.no_grad():
    out = model.generate(**inputs, max_new_tokens=256, temperature=0.2)
print(tokenizer.decode(out[0], skip_special_tokens=True))

Training

Stage 1: SFT

Started from Qwen/Qwen2.5-3B-Instruct
Same recipe as the 1.5B run: scripts/train_sft.py with oracle JSON-action trajectories, TRL SFTTrainer, LoRA (defaults such as LORA_R=16, NUM_EPOCHS=3, eight seeds over the train scenario split — see the incident-triage-env repo)
Compute: 3B was trained on A10G (the 1.5B reference run fit a T4 in ~218 s; larger bases use the same code path with more VRAM)

Stage 2 (optional): GRPO refinement

Optional TRL GRPOTrainer refinement on top of this adapter is supported by the same codebase as the 1.5B pipeline. This Hub upload is the SFT checkpoint unless you document a separate GRPO-tagged revision.

Evaluation summary (held-out hard tasks)

Use the same Phase 8 harness and three held-out tasks as the 1.5B model card: hard_multi_signal_cascade, expert_stealth_regression, and hard_pr_quality_breach, compared against baseline-hf, reporting composite and per-head scores.

bash

# From the incident-triage-env repo, with this adapter + base wired in eval config:
python scripts/eval_before_after.py --compare baseline finetuned

Fill in quantitative deltas for this checkpoint from your run logs (they are not duplicated here so the card stays honest to per-size measurements).

Limitations and risks

still sensitive to prompt format and missing context
diagnosis can lag oracle-style references even when blast/PR heads are strong
may produce over-confident recommendations under ambiguous telemetry
should always be reviewed by an on-call engineer

Recommended Hugging Face metadata (UI fields)

base_model: Qwen/Qwen2.5-3B-Instruct
pipeline_tag: text-generation
library_name: peft
license: apache-2.0
language: en
tags: incident-triage, sft, lora, peft, openenv
datasets: your SFT training dataset identifier(s) (oracle trajectory bundle from the env)
metrics: composite_score, optionally per-head metrics (diagnosis, policy, blast, pr)

Model card contact

Open an issue in the incident-triage-env / hackathon repository for bug reports, misuse reports, or benchmark reproduction requests.