NextToken-model-2 — SAM (Small Action Model)

A 272.7M-parameter Hindi-English model fine-tuned into a small action model for banking/NBFC customer service: given a user utterance and a system prompt listing the available tools, it decides whether one applies and extracts the arguments.

This checkpoint (sep19_round1, end of epoch 1) replaces sep18_round2. Updated in place — this repository tracks the current best checkpoint; check back for updates.

What changed vs. sep18_round2

Same 5-tool + no-tool scope, but a substantially larger and more categorized training set: 15,894 examples (14,196 train / 1,698 val), organized into 8 explicit buckets — 5 tools, negative (with 4 sub-categories: chitchat, definitional, info, statement), no_tool, and a general_capability bucket (non-banking text: continuation, summarization, lists, creative writing) specifically to keep the model useful outside the banking domain and reduce over-firing a tool on unrelated prompts.

Early-stopped again, same lesson as before: trained to epoch 1 (this checkpoint), then epoch 2 showed the same overfitting signature already seen in this project — val loss rose (1.5426 → 1.6514) while train loss kept falling — so training was stopped before epoch 3. The published weights are the end-of-epoch-1 checkpoint, not a later one.

Verified accuracy — held-out 80-example eval, by category

This is the training team's own held-out evaluation (eval_epoch1_results.json, never trained on), broken down here rather than reported as one aggregate number:

negative_statement is a real, systematic failure, not noise: all 8 examples in this bucket are past-tense/statement framings (e.g. "just checking in about my home loan"), and the model fired the identical get_loan_balance/get_payment_status-style tool call on every single one instead of giving a natural-language acknowledgment — the same "inconsistent negative-suppression for past-tense/third-person phrasing" limitation flagged on the previous checkpoint's card, now precisely quantified rather than described qualitatively.

Independently re-verified here (not just taking the training team's numbers on faith): re-ran two real training examples (one get_loan_balance positive, one negative_info) through the converted HF checkpoint — both reproduced their exact target string.

Architecture

Same base architecture as somasekhar-dev/NextToken-model-1:

Usage

Tokenize each message segment separately and concatenate — do not re-tokenize a merged prompt string (a real train/inference tokenization mismatch was found and fixed earlier in this project's history):

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model = AutoModelForCausalLM.from_pretrained("somasekhar-dev/NextToken-model-2")
tokenizer = AutoTokenizer.from_pretrained("somasekhar-dev/NextToken-model-2")

system_prompt = (
    "You are a banking assistant for an Indian NBFC. Today's date is 2026-09-19 (Saturday).\n"
    "Given the user's message, decide whether one of the tools below applies. If it does, "
    "call it: reply with a short natural-language line if appropriate, followed by a JSON "
    "object {\"tool\": \"<name>\", \"arguments\": {...}} on its own line. Only use arguments "
    "that are stated in the conversation or can be computed from today's date -- never invent "
    "a value. If no tool applies, reply in natural language only, or with {\"tool\": null} if "
    "a bare JSON response is expected.\n\n"
    "Available tools:\n"
    '{"name": "get_loan_balance", "description": "Outstanding principal on a loan", '
    '"parameters": {"type": "object", "properties": {"loan_type": {"type": "string", '
    '"enum": ["home", "car", "personal", "gold", "business"]}}, "required": []}}'
)
user_utterance = "मेरे लोन का कितना बकाया है, कृपया मुझे बताएं।"

segments = [f"System: {system_prompt}\n", f"User: {user_utterance}\n", "Assistant: "]
ids = []
for seg in segments:
    ids.extend(tokenizer.encode(seg, add_special_tokens=False))
ids_t = torch.tensor([ids])

out = model.generate(ids_t, max_new_tokens=60, do_sample=False,
                      eos_token_id=tokenizer.eos_token_id, pad_token_id=tokenizer.pad_token_id)
print(tokenizer.decode(out[0][ids_t.shape[1]:], skip_special_tokens=True))
# -> {"tool": "get_loan_balance", "arguments": {}}

Training

Full fine-tune (not LoRA), 15,894 examples (kashyap/sep19/raw/*.jsonl), 8 buckets (5 tools, negative x4 sub-types, no_tool, general_capability), val loss 4.8757 -> 1.5426 by end of epoch 1. Stopped before epoch 3 on detecting overfitting (see above).

Status and limitations

• negative_statement framing is unreliable — 0% on the held-out eval. Do not rely on this checkpoint to correctly ignore past-tense/statement mentions of a loan or account; it is likely to fire a tool call anyway.
• negative_chitchat is inconsistent (62.5%) — casual conversational openers sometimes still trigger a tool call.
• Tool scope is deliberately narrow: 5 tools + no-tool. The full 41-tool schema (model2_tools.json) does not fit this model's 2048-token context at all (3,359 tokens alone) — this is an architectural constraint, not just a training choice.
• Free-text argument values and multi-turn slot-filling are not covered by this round's eval — treat as unverified until tested.
• Any state-changing action should be confirmed before executing, not trusted blindly from a single model output.
• Conversion to this Qwen3ForCausalLM checkpoint was verified numerically (100% argmax agreement vs. the original training-format checkpoint on random inputs, a full save/reload round-trip, and two real training examples reproduced exactly).